Wolf Digest — Wednesday, August 26, 2026

The Royal Netherlands Navy will equip its fleet with a kinetic counter-drone capability built on Destinus's Hornet B2 interceptor, with delivery, integration onto several warships and support toward an initial operating capability scheduled to complete in the second quarter of 2027. The system pairs a base station with a launch platform; the interceptor leaves a sealed canister on a rocket booster, deploys wings and transitions to electric cruise flight, then uses electro-optical and infrared sensing with onboard AI processing for terminal guidance. The Dutch Ministry of Defence framed the requirement around medium-range drone threats identified in both operational analysis and the current maritime threat picture. The earlier Hornet B1 is already in serial production.

The economics are the point. Western navies have spent the Red Sea campaign firing multimillion-euro missiles at cheap Houthi drones, and Ukraine has demonstrated low-cost interceptor drones against one-way attack UAVs at scale. The U.S. Navy has already installed Coyote interceptor launchers on Arleigh Burke destroyers. Destinus puts cost per engagement far below traditional interceptor missiles, and its chief executive framed the requirement as inverted air-defense economics that no navy can sustain by trading expensive missiles against volume.

Naval Air Systems Command issued a request for information for the Silent Anvil Air-Launched Torpedo, a standoff anti-submarine weapon designed to be captive-carried and released from internal and external aircraft weapons stations using standard carriage systems, then to operate as a precise, highly maneuverable glide weapon that transitions quietly to the target area. The template release altitude band runs from 10 to 60 kilofeet, and the Navy is asking industry what maximum glide range is achievable, with a performance goal of maximizing aerial flight endurance while minimizing time in the water.

The form factor is specified to resemble the GBU-39 Small Diameter Bomb with BRU-61/A bomb rack compatibility, explicitly so that the widest possible set of manned and unmanned platforms can carry it, and the RFI notes the Air Force could also use the weapon. The payload combines an acoustic sensor, warhead, command-and-control section and electric propulsion, with a wing kit, radio and aperture, and guidance and control on the flight side. The program intends to use government-owned and standard software interfaces for fast, inexpensive reconfiguration, and requires integration with tactical command-and-control tools for coordinated operations and in-flight target updates. The stated goal is expanding the range of drones and manned aircraft that can conduct anti-submarine operations.

Generalist has raised close to 200 million dollars led by 8VC at a 3 billion dollar valuation, according to a regulatory filing and two people with knowledge of the round. The money extends the 400 million dollar Series B that Radical Ventures led in June at a 2 billion dollar valuation, bringing the round to 600 million total. The company was founded in 2024 by former Google DeepMind researchers Pete Florence and Andy Zeng with former Boston Dynamics engineer Andrew Barry, with early backing from 8VC, Radical Ventures, Nvidia, Union Square Ventures, Bezos Expeditions and Fei-Fei Li.

Generalist is building a robot foundation model intended to work across hardware, and claims its newly released Gen 1.5 model lets robots acquire new tasks from video demonstrations as short as three to twelve seconds. It is working with a small set of customers and tuning the model against their feedback. The comparison set is crowded and expensive: Physical Intelligence is reportedly valued at 11 billion dollars, SoftBank-backed Skild AI at 14 billion, and Genesis AI was in talks last month at 3 billion. TechCrunch notes the standing objection to the whole category, which is that robots cannot be trained on the entirety of the internet the way language models can, so a genuinely general robot policy may still be years out.

Army Network Enterprise Technology Command transferred network and command-and-control services between geographic theaters for the first time in the run-up to the Iran conflict, a capability the service had planned for but never executed operationally. The move matters because Army network services have historically been tied to the theater that owns them, which limits how quickly signal capacity can follow a shifting main effort. Doing the transfer under real operational pressure rather than in exercise conditions is the datapoint, and it feeds directly into the Army's broader Next Generation Command and Control effort and its stated ambition to run network defense with agentic tooling rather than analyst headcount.

VivSoft Technologies received a 100 million dollar Other Transaction Authority production agreement for the Aerospace Readiness Enterprise System, which will fold six separate Air Force scheduling, training and readiness platforms into a single software-as-a-service environment inside the service's new Operations Enterprise Architecture. Rollout across major commands begins in 2027. The award followed a 90-day competitive prototype phase in which multiple vendors built versions of ARES and handed them to operators for testing, and because the winning prototype was already evaluated by airmen the program transitions straight into production. The current patchwork requires manual data entry and gives commands little visibility across each other; ARES replaces that with automated scheduling, training and readiness management, including assignment of personnel and resources and automated conflict detection.

The General Services Administration's Login.gov is looking at expanding device fingerprinting as a fraud control, with automated account creation by bots and AI agents named explicitly alongside conventional fraud as the threat being addressed. It is an early instance of a federal identity system having to decide how it treats an agent acting on a legitimate user's behalf, which is the same question every consumer platform is now facing but with statutory access obligations attached. The tension is that fingerprinting is a privacy-sensitive control being proposed for a service whose users cannot opt out of interacting with government.

A CSIS Strategic Technologies commentary frames the shift from petrodollar to what it calls tokendollar dynamics, arguing that the leverage the United States historically drew from denominating energy trade in dollars is being reconstituted around AI compute and inference capacity. The piece reached the digest through the Google News proxy rather than directly, so this summary is built from the feed metadata alone and does not reflect the full argument. The framing is worth flagging because it sits alongside two other items today, the compute-concentration thesis and two hyperscaler accelerator disclosures, that all point at inference capacity becoming the scarce strategic good.

Recuris addresses the failure mode where a long-horizon agent's growing history obscures the current task state and misaligns skill invocation. Working Memory tracks task progress and drives skill selection from Experiential Memory, so skills are chosen against present need rather than the full trajectory. That coupling also makes execution into structured evidence that localizes failures to specific memory components, which a fixed Meta-Agent converts into validation-gated local updates to Skill Memory, closing a bounded recursive loop. Across four long-horizon benchmarks and ten models, Recuris improves success in 35 of 37 completed model-benchmark pairs, adding 17.8 points to GPT-5.6 Sol and 15.6 to Claude Opus on tau-bench.

Self-improving agents refine answers rather than the process producing them, and systems that add a meta-level either freeze that level or must leave part of their own editing machinery untouched to stay stable, which caps realized meta-depth at roughly two. Meta^n instead holds the meta-operation Omega fixed and recurses on its input: Omega reads the traces of the solver stack below along with the code that produced them, and writes the next layer as a strategic pre-process plus a library of callable helpers. Because Omega never changes it cannot destabilize the system, and because its input strictly grows each layer reasons from a higher vantage. Depth is set by convergence rather than fixed in advance, with an evolutionary archive searching over layer chains. Meta^n beats prior self-improving agents on all eight benchmark families across two backbones.

Automated parking research usually assumes marked slots and short approach maneuvers, whereas delivery and service vehicles have to reach an operator-specified pose in an irregular bounded environment from a distant start. Learning-based planners that rely on local observations lose the long-range route context that decides those approaches. NeuralParker encodes obstacle and boundary geometry for the whole environment in a target-relative vertex representation, then couples a learned curvature-length arc policy with an in-loop terminal ensemble that picks among cubic Hermite connections under a curvature-regularized cost. The authors also contribute factorial and long-range route-choice benchmarks, since existing ones do not test the failure mode.

Action-conditioned world models are increasingly used as learned simulators for policy evaluation and improvement, which rests on an assumption nobody had verified: that generated futures faithfully reflect arbitrary valid actions, not just the ones in the demonstration set. Existing benchmarks stay inside expert demonstrations. WorldEcho probes action following over a broader action distribution using visual integrity and SE(3) trajectory alignment, and finds current world models execute expert actions reasonably but struggle off-expert, which is precisely the regime that matters if the model is being used to improve a policy rather than replay one.

RAND's Center on AI, Security, and Technology has published Security Level 3, a standardized framework of 262 security controls adapted from NIST Special Publication 800-53 and aimed specifically at protecting AI model weights from organized cybercrime and insider threats. The controls are mapped to 31 attack vectors rated high-feasibility, and the framework is explicitly designed for incremental adoption within a six-to-twelve-month window rather than as an all-or-nothing posture. It is a direct successor to RAND's 2024 Securing AI Model Weights work, which defined the security-level ladder but stopped short of an implementable control set. The practical significance is that labs and their government interlocutors now have a concrete artifact to point at when arguing about what a given security level actually requires.

Chris Malone, who oversaw execution of OpenAI's data center strategy, left the company last week after joining in March of last year, following nearly five years at Meta and more than a decade at Google. OpenAI described a recent reorganization of its infrastructure organization; as part of it Malone had stopped reporting to president Greg Brockman and begun reporting to vice president Sachin Katti. The company named Uday Ruddarraju, Brent Mayo and Spas Lazarov as the executives now covering the data center team, build and delivery program, and engineering respectively.

The departure is notable mainly for its seat. Data center strategy is arguably the most closely watched role at any lab right now, and Malone joined shortly after the Stargate Project launched. Business Insider counts thirteen executive departures at OpenAI in 2026, including chief revenue officer Denise Dresser after eight months, longtime chief operating officer Brad Lightcap, and Fidji Simo, who stepped back for health reasons. The safety side has also thinned, with the head of ethics leaving in July and the preparedness team reported disbanded last week. Brockman has argued the scrutiny is a function of the spotlight rather than the churn rate, but the timing sits awkwardly against a public listing now reportedly pushed to 2027.

A DHS inspector general report found that Homeland Security Investigations' technology base materially hindered its efforts to counter fentanyl trafficking, with agents describing the tooling in blunt terms. The finding is a reminder of the gap between the analytic capabilities being procured at the policy layer and what investigators actually have on their desks: data that cannot be joined across systems, case management that does not surface links, and no practical path to applying pattern-detection tooling to the underlying records. It is the unglamorous half of the federal AI story, where the binding constraint is data infrastructure rather than model capability.

Shipping models that are both structurally compressed to a fraction of their parameters and quantized to 4 bits degrades reasoning, mathematics, coding and long-context behavior enough to need a recovery stage. The default answer, quantization-aware training, re-fits the compressed and quantized model to hard labels; in this pipeline it converged slowly and collapsed past its peak. Quantization-Aware Healing instead distills the 4-bit student directly from the original uncompressed model, on the argument that a structurally compressed model's bfloat16 checkpoint was itself only a distillation-recovered approximation, so hard labels are the wrong target. The result reported on a GPT-OSS-class model is a compressed 4-bit checkpoint that outperforms its full-precision original.

Latent world models are strong encoders, but action planning on top of them still runs an iterative optimizer from scratch for every state-goal pair, treating the world model as a black-box simulator and paying the full optimization cost anew at each replan with no reuse of planning experience across queries. LeFlow asks whether planning itself can be amortized once the latent world model is learned, replacing per-query trajectory optimization with a generative latent flow over plans.

Robot policies take heterogeneous observations at each decision step, and sequence models differ in how they organize those inputs over time. WorldToken is a time-first instantiation that fuses multiview images, proprioception and task conditioning within each policy timestep into a single world token, models the resulting sequence with a causal temporal Transformer, and generates action chunks with a diffusion head. On 23 RoboCasa tasks an 85.3M-parameter policy trained from scratch apart from a frozen CLIP text encoder reaches 59.45 percent mean closed-loop success from 2,900 demonstrations, which is a notable parameter budget for that result.

The Treasury Department is establishing a task force to help the financial sector plan for migration to post-quantum cryptography. The relevance to this digest is indirect but real: the harvest-now-decrypt-later threat model and the AI model-weight-exfiltration threat model share the same institutional response, which is a long, expensive inventory-and-migrate program that has to be started well before the capability it defends against exists. Treasury moving on the cryptographic version while RAND publishes a 262-control weight-security standard on the same day is a reasonable illustration of how slowly this class of work moves relative to the capability curve.

Endpoint rewards tell a diffusion model that a sample was good or bad but not how any intermediate denoising prediction should change. DiffusionOPSD has a frozen behavior policy generate trajectories that supply query states and anchors, then uses reward gradients to construct bounded positive and negative targets around each anchor, which the trainable policy fits as detached supervision before an exponential moving average refreshes the behavior policy. Separating target construction from finite realization lets the authors measure each; controlled same-query experiments show larger target-construction gains do not reliably translate into larger realized gains after one fitting update, which is a useful negative result for anyone assuming reward-gradient magnitude is the thing to optimize.

Reinforcement learning with verifiable rewards gives sparse task-level feedback; on-policy distillation gives dense token-level guidance but ignores trajectory correctness and so caps the student at the teacher. Existing combinations rely on weighted mixtures or heuristic switching, which introduce hyperparameters and trade-offs. OPDVR reformulates the implicit reward of sampled-token on-policy distillation in terms of trajectory correctness, then applies a ReLU gate so correct trajectories receive non-negative reward and incorrect ones do not, which merges the two objectives without adding a single tuning knob.

LAION-BVD collects 1.3 billion platform-specific video URLs from CommonCrawl and downloads 80 million of them for 10 million hours of footage, targeted at joint video, audio and image pretraining. Content-aware scene detection cuts clips, and video and audio captions are generated synthetically. Models trained on the corpus are competitive on standard video-text and audio-text benchmarks with consistent gains as training or model scale increases. The authors also extract scene-changing frames as an alternative image-text source and note that their visual distribution differs measurably from standard web image corpora, with strong image-text retrieval results. The dataset is released openly, which is the main contribution given how little open video data exists at this scale.

Terminal rewards neither localize intermediate retrieval errors nor redirect a trajectory before those errors compound. CAFE has a shared-parameter model alternate between search-agent and critic roles, so the agent learns when to request and use feedback while the critic learns to infer corrections from outcome-confounded rollouts whose failure patterns shift as the agent improves. Feedback-conditioned recovery is initialized from the base agent's own failures, then online and offline optimization are coupled: a comparative feedback estimate uses the prompt-level call-versus-skip success gap to shape request returns, and feedback-aware advantage shaping reweights token advantages before and after the intervention.

EchoWM organizes interaction around camera intent: in first-person scenes the camera specifies observer motion, and in third-person scenes camera-character dynamics are learned from data with no view-specific controller. Discrete commands and continuous poses map to a shared metric-scale relative 6-DoF trajectory, with dataset-level calibration preserving motion magnitude across heterogeneous sources. Training combines a complementary data engine with a progressive schedule followed by autoregressive post-training for long-horizon generation. The model reports strong trajectory following and visual quality on public world-model benchmarks while jointly producing video, environmental sound, music and speech, which is a wider output surface than most enterable-media systems attempt.

The paper formalizes prefix invariance, the requirement that a representation at position t must not depend on future inputs, and gives a lightweight audit needing two forward passes, no training and no gradients, that localizes exactly where causality breaks. The key observation is that inspecting the attention mask is not sufficient, because leaks can enter through scans or normalization even when the mask is correct. Across 192 injected-fault trials on eight checkpoints, mask inspection caught none and the audit localized all 192, and the same procedure surfaced real defects in Zamba2 and Nemotron-H. For anyone shipping hybrid attention-plus-state-space stacks this is a cheap check worth running.

Agentic systems increasingly gate actions on the model's own stated confidence, which assumes confidence tracks correctness at the moment of acting. The authors test that in a chess variant where royal status can be secretly and repeatedly relocated between pieces, eliciting a probability distribution over the opponent's hidden royal piece every turn separately from the chosen move, and scoring it against ground truth recoverable after the game. Captures made at stated confidence of 0.5 or above were correct in 1 of 62 cases. The calibration deficit is almost entirely concentrated in exactly those high-confidence action events: 99.3 percent of it in the original batch and 98.7 percent in the replication. The ordering holds in weaker form across four further model configurations spanning a second provider.

The locate-then-tune paradigm isolates critical parameters interpretively and then restricts fine-tuning to them, but mechanistic interpretability is retrospective, so interpreting the pre-SFT model gives misleading answers: on novel tasks the neurons identified up front differ sharply from those that end up governing the tuned model, and the resulting bias actively disrupts training. The authors model supervised fine-tuning as continuous parameter evolution and use a Taylor expansion to connect the post-tuning mechanistic objective to the pre-SFT model's dynamic gradients, giving a dual-granularity estimate of the post-SFT interpretability state from pre-SFT parameters and the target dataset alone.

Gameplay video is an appealing source of world-model training data, but raw footage entangles the game world with screen-space interfaces, injecting game-specific bias and irrelevant dynamics. G2WEngine formalizes user-interface grounding and removal under a GameUI-Taxonomy, automatically extracting reusable UI assets from real footage and synthesizing temporally coherent overlays onto clean video. The resulting Game2World corpus has 96,000 synthetic paired videos with exact reconstruction targets plus 1,079 in-the-wild clips from 303 games for realistic evaluation, backed by an asset library of 5,132 verified UI elements across 21 categories.

Reinforcement post-training for video multimodal models is sample-inefficient because on-policy groups contain few high-quality rollouts even with expensive chain-of-thought generation. OraRL's observation is that an annotation can enter the on-policy group as an oracle rollout and act as a direct positive optimization target rather than merely scoring other rollouts. Naive integration breaks, because a high-reward oracle raises the group baseline and inverts otherwise-positive policy advantages, a failure the authors name advantage inversion. The fix is a decoupled advantage estimator in which policy rollouts set an oracle-free baseline while the oracle-policy gap modulates a directional gain and a separate detached oracle term.

Most supervised machine-generated-text detectors degrade out of domain and need large diverse training sets. This work shows the latent representations of generated and human text are linearly separable in low dimensions, attributes the separability to systematic differences in representation quality, and trains two variants of a plain linear probe. Across four benchmarks against sixteen baselines the probes improve out-of-domain AUC by eleven points while reaching near-peak performance from fewer than 100 samples. The transferability argument is that probes recover a shared latent direction rather than domain-specific surface features.

Mamba-style state-space models have been applied to video anomaly detection, but existing systems still buffer clips or windows internally, offer no theory linking temporal memory to detection latency, and benchmark only GPU throughput rather than the edge hardware they target. This detector keeps a fixed-size state updated in constant time and memory per frame with no lookahead and no clip buffering, using a diagonal linear state-space recurrence with an input- and state-dependent decay gate trained self-supervised through causal next-embedding prediction on a frozen backbone. The contribution is a closed-form relationship between the recurrence decay spectrum and both detection delay and the shortest reliably detectable anomaly, validated against the settling-delay bound on UCSD Ped2 and CUHK Avenue.

WeChat's universal multimodal embedding family covers text, images, video, visual documents and arbitrarily interleaved inputs with flexible output dimensions, in 2B, 4B and 9B variants. Training runs in two stages: large-scale multimodal alignment, then refinement with curated data, fine-grained relevance supervision and cross-scale knowledge transfer. The 2B variant already exceeds the previously leading 8B open-source baseline on MMEB-v2, and the 9B sets a new state-of-the-art overall score of 80.6, which is the more useful datapoint for anyone budgeting embedding inference at scale.

Coding agents re-send large file reads and tool outputs every turn, and that context dominates the token bill. General-purpose prompt compressors trained on prose paraphrase identifiers and drop the exact spans an agent needs to edit. Paritok-4B is a 4B LoRA compressor built on two commitments: it is extractive, selecting spans rather than rewriting them, with 96.0 percent of emitted identifiers, paths and numbers already present in its input and 96.2 percent on held-out SWE-bench Lite; and it is intent-conditioned, acting chiefly inside a retained segment to decide which lines survive rather than changing how much is retained. It distils a gpt-4.1-mini teacher over 67,074 real OpenHands trajectories into 40,606 validated examples on top of Qwen3-4B.

Most EEG foundation models are reconstruction-centered, which implicitly assumes that signal predictable from local context is transferable neural information. INCEPT instead trains for representation-level stability across correlated EEG observations, separating stable neural structure and essential subject-sensitive information from the nuisance variability that dominates scalp recordings, while preserving subject-, state- and condition-discriminative content. Pretraining runs on over 11,000 hours of unlabelled clinical EEG, and evaluation spans ten datasets covering signal-level assessment, brain-state decoding and brain-health tasks.

Ockhamareto is a single-shot GRPO framework for unit-test generation combining a Pareto-gated bonus that rewards only rollouts non-dominated in mutation-score versus test-count space, and token-level segment credit that attributes each test's marginal mutation kills back to the tokens of its own test block. On UnLeakedTestBench it strictly Pareto-dominates MIST-RL and wins on every objective separately, catching more bugs at 49.9 percent mutation score versus 31.3 at N equals 5 while using 2.60 tests on average against 4.67, for 3.4 times the per-test trade-off improvement. The advantage holds across HumanEval+, MBPP+, CodeContests and TestGenEval-Lite.

Live commerce distributes product facts across speech, video frames, product images, overlaid text and user queries in noisy, temporally extended streams. TLive-Omni maps image, video, audio and text into one representation space and introduces Per-vGrid, a timestamped token organization that groups each video grid with its temporally corresponding audio inside explicit boundary tokens to force temporal alignment. Training runs three supervised stages from omni-modal perception to instruction following, then a Faithful-RFT reinforcement stage that scores final responses with task verifiers to improve faithfulness and expression quality under real-time constraints.

A CSIS assessment of how fast the Defense Department could rebuild and reconfigure munitions production capacity. The item reached the digest through the Google News proxy, so this note is written from feed metadata rather than the full text. The relevance here is the same throughput-versus-demand framing that runs through today's counter-drone and standoff-weapon items: the constraint on both is production capacity and per-engagement cost, not sensing.

Improving an LLM judge normally requires annotations, reward models or distillation from a stronger teacher. RecurSE closes the loop: a trainable judge evaluates candidate responses under per-rule rubrics, and a synchronized policy-copy checker audits the judge's reasoning against meta-rubrics to supply a scalar process reward. Interface decoupling structurally isolates the checker's score from the judge's verdict tokens, which is what eliminates the degenerate token-copying shortcut that otherwise collapses this setup. The paper is explicitly framed around two questions worth asking of any self-improvement scheme: when can it occur, and when must it stop.

A trustworthy spreadsheet analysis needs a valid path from the question to the supporting evidence, which yields two diagnostic questions: can the model refuse or ask for clarification when no such path exists, and does it preserve the correct analysis when the same evidence is re-expressed in a different table shape. TrustDABench derives 19 perturbation operators from that evidence-path view and instantiates them through an agentic generation framework, producing 2,340 human-verified perturbed instances. Across eight models the best reliability result is 24.21 percent average, which leaves the headroom unusually visible.

Massively parallel simulation changes the data regime that off-policy reinforcement learning trains in, and the stabilizers designed for data-limited replay do not transfer. Controlled experiments across eight benchmark families show parameter normalization helps under narrow replay coverage but restricts value fitting when data are abundant, and clipped double-Q can be relaxed in high-throughput manipulation, while age-biased replay weighting helps across regimes and especially with limited network capacity. WarpSAC packages this as two variants: normalization on with clipped double-Q for CPU-scale data-limited training, and normalization off with a single Q for GPU-parallel data-abundant training, improving normalized-score-per-step AUC over FlashSAC by 4.5 percent across nine tasks.

Smart glasses are moving from capture-and-display accessories toward first-person intelligence platforms that connect human perception, persistent context and digital or physical action, with an on-body viewpoint aligned to the wearer's vision, hearing, motion and hand-object interaction, under tight energy, thermal, privacy and feedback constraints. The literature is fragmented across devices, tasks and benchmarks. This survey argues the question is not whether a model can recognize, answer, remember or act in isolation, but whether a complete system sustains a reliable, temporally valid, correctable and governable loop, and formalizes first-person data flow and constrained task utility to make that comparable.

Tool-augmented models are bounded by the APIs someone bothered to write, and existing tool-creation systems patch that by prompting a frozen model at inference time, leaving the model that writes a tool decoupled from the one that uses it with no signal that the schemas it emits are schemas it can invoke. SMITH makes each rollout either a build task, writing a tool from a few examples, or a use task, invoking a pooled tool on a held-out question, with three separate reward axes catching schema, code and outcome failures independently so each failure mode contributes its own gradient. A 4B Qwen3 trained this way across 13 procedural reasoning tasks is the reported result.

Group-relative reinforcement learning waits for sibling rollouts of the same prompt, which is expensive for long variable-length tool-use trajectories. Single-stream Policy Optimization removed that dependency with a persistent prompt-level value estimate, but whitens one advantage per trajectory before optimizing a token-mean actor loss. The authors show trajectory centering generally does not center the token-weighted quantity the actor actually consumes, and fix it by standardizing terminal-outcome advantages under the action-token measure, while organizing prompt evidence by the policy event that generated it rather than by learner receipt order. Matched runs on ALFWorld at two scales and on Math-TIR show improved online learning efficiency over SPO.

Safeguarding agents requires assessing complete execution trajectories under context-dependent policies, and prompting or supervised fine-tuning limits adaptation to unseen trajectories and changing policy contexts. RePolicy takes a trajectory plus a dynamic policy library, invokes the applicable policy, and uses its content to produce a policy-grounded rationale and judgment. Training uses PolicyTraj-20K for supervised initialization then GRPO with verifiable rewards and policy-context perturbation, evaluated across six agent-safety benchmarks.

Ternary transformers give extreme memory and compute efficiency, but low-bit LoRA methods cannot fine-tune ternary weights directly: they either dequantize to higher precision to merge adaptation weights, or update only quantization parameters and never produce a merged model that stays ternary. This work represents discrete updates such as sign flips and zeroing through a low-rank Kronecker factorization into two small ternary matrices applied elementwise, which stays inside the ternary domain and supports direct merging. Experiments cover six models across language and vision.

Procedural video-language models solve heterogeneous tasks from the same visual evidence, and dense decoders share feed-forward networks across all of them, entangling task behavior and making controlled capability expansion hard. Sparse mixture-of-experts decoders give conditional computation but token-level learned routing does not align with task-level procedural objectives. MoTE turns feed-forward networks into task-specific experts while keeping the multimodal backbone shared, with one sample-level task route per example, so active expert computation stays independent of how many task experts are stored. The five-expert VideoLLM-MoTE activates roughly 2B language-model parameters per example across five COIN benchmarks.

Claims about when concepts emerge during training and whether linear probes recover real structure are hard to validate in language models, because language supplies no ground-truth ordering of concepts or relationships among them. The authors propose using astronomy instead, probing AstroPT, a transformer trained on millions of galaxy images, across checkpoints, layers, model sizes and objectives. Galaxy properties emerge in a fixed order that tracks their known difficulty, with quantities written almost directly into the pixels appearing first, which gives probing methods a calibration target they otherwise lack.

Repository understanding methods lean on surface-level code retrieval and cannot reason across multiple files, complex architectures or long-range dependencies. DeepRepoQA puts agents inside a systematic tree search over repository structure, using Monte-Carlo tree search to decide what to navigate and inspect next, which turns multi-hop dependency reasoning into a search problem rather than a retrieval one. Results on SWE-QA report substantial gains over strong baselines.

Financial tabular data carries domain constraints, severe class imbalance and asymmetric attacker capability, which makes robustness evaluation unusually protocol-sensitive. FraudBench evaluates the same dataset-model-attack-defence setting under three matched protocols — unconstrained attacks, post-hoc feasibility filtering, and deployment-aware constraint-integrated attacks — across four public financial datasets, and shows the three can support materially different conclusions about the same system.

Earthquake early warning depends on rapid magnitude estimation, and operational systems typically need dense regional networks and region-specific calibration, which leaves sparsely instrumented high-risk areas uncovered. Single-station learning is cheaper but existing models trade accuracy against latency and degrade under regional distribution shift. SeisMamba combines hierarchical convolutional encoding, sparse selective state-space modelling, multi-scale feature fusion and an auxiliary temporal prediction head over minimally processed three-component waveforms, achieving the best MSE, RMSE and R-squared among tested baselines on STEAD at 0.55 milliseconds for a batch of 32.

Combining computational text analysis of millions of AI-related newspaper articles and social media posts with 57 semi-structured interviews of AI professionals conducted in 2021 and 2023, before and after the surge in public interest, the authors identify the interpretive frames professionals use to handle cognitive challenges such as assigning responsibility for societal impacts. They organize those frames across three debates: the method of AI development, between top-down expert systems and bottom-up emergent capabilities; the mind of a system, ranging from passive tool to humanlike digital mind; and the morality of the enterprise.

Agent benchmarks mostly evaluate tool selection, argument generation and end-to-end success under serial execution, leaving valid parallelization and resource-constrained scheduling untested. That creates a real failure mode: serial execution is safe but slow, resource-agnostic parallel execution is fast but overflows budgets. PeakBench provides executable multi-tool workflows with execution-grounded dependency annotations and measured resource profiles, and addresses attribution so a failure can be traced to scheduling rather than to selection or arguments.

RL-trained agents remain episodic and accumulate no reusable knowledge across episodes. Skill-based approaches such as SkillRL extract skills from raw trajectories but treat the skill bank as append-only, never checking whether a stored skill still works. SkillForge makes skill usage explicit during interaction so reinforcement learning can optimize environment actions and skill invocation decisions jointly, and adds evidence-based skill verification so skills are refined or retired rather than merely accumulated.

Process-level factual supervision is the standard mitigation for hallucination risk under outcome-driven rewards, but coarse aggregation of factual signals and the absence of any reliability assessment for them create a mismatch between what gets verified and what gets updated. The authors decompose this noisy factual credit assignment into credit localization ambiguity and credit reliability ambiguity, and FARCA addresses both by mapping fact-level verification onto specific tokens and weighting each signal by its assessed reliability.

Policy optimization for language models mediates the stability-exploration trade-off with an action-side policy-KL regularizer, which puts practitioners in a bind: keeping it constrains response behavior and consumes the action-side exploration budget, dropping it removes explicit drift control. ERPO argues for regularizing the input side instead, adding a query-KL term that bounds how far the policy-induced distribution over training queries drifts from its pre-RL reference, together with a dataset-static per-query weight derived from that reference.

In multi-turn service agents the task information unfolds over time as users clarify goals and tools return data, so a final reward cannot indicate which actions resolved the task. Existing methods lean on comparative evidence from other trajectories, resampled continuations, or separately constructed step-level signals. IAPO's observation is that a completed rollout already records how information and errors propagate between agent actions, and that influence structure can be used directly for credit assignment.

Marginalized importance weighting reweights offline samples by the target policy's discounted occupancy ratio, characterized by an adjoint Bellman equation, but minimax, primal-dual and fitted fixed-point estimators all leave residual occupancy-balance violations from function-class approximation, regularization or incomplete optimization. Those violations are hard to diagnose because the objectives lack a direct supervised validation loss for tuning, model selection or early stopping. Isotonic Bellman calibration is a one-dimensional model-agnostic post-processing step that reduces the violations while preserving the ranking information in any initial estimate.

Test-time reasoning has improved problem-solving but autoregressive decoding still executes long traces sequentially, producing latencies measured in days for the hardest tasks. Prior parallel-reasoning systems focus on subtask parallelism, decomposing a task into independently solvable chunks, and overlook trial parallelism, where multiple speculative attempts explore and verify concurrently. Parason separates the two and shows both are pervasive in real reasoning traces.

Full-length RNAs, especially messenger RNAs, routinely exceed the context lengths used to pretrain existing RNA foundation models, which prevents complete-transcript modeling at single-nucleotide resolution. RIBOSPAN combines dense bidirectional self-attention, single-nucleotide tokenization and attention-isolated sequence packing to pretrain natively at up to 10,240 nucleotides, and is evaluated through nucleotide reconstruction and controlled long-context representation tasks.

Precision oncology needs a longitudinal patient-state model that captures disease evolution and treatment over time across modalities. The oFM was developed on a real-world cohort of 1.67 million cancer patients, with over a million reserved for training under patient-level partitioning, integrating daily clinical and molecular episodes with DNA, RNA and hematoxylin-and-eosin pathology images into a time-integrated patient-state representation.

Finding stable neuron behavior across a whole domain is hard because existing methods use instance-level point estimates that obscure population variability, or procedures too expensive for domain-wide analysis. RACE is a forward-pass statistical framework using residual alignment to evaluate domain-wide functional consistency of transformer neurons. Perturbation experiments show better domain specificity than gradient-based point estimates, token-distribution results confirm the selected neurons associate with the target domain, and the compute overhead is roughly a hundredth of gradient approaches.

Evaluation normally assumes observable behavior is determined by weights, training data, alignment procedures and prompts. The paper argues that is incomplete, because modern inference pipelines can systematically modify the probability distribution immediately before token selection, creating a control layer between frozen weights and observed text. Controlled generation work such as PPLM, GeDi, DExperts and FUDGE, and watermarking systems such as SynthID-Text, establish that decoding- and logit-level intervention is technically mature; the contribution is formalizing what undisclosed use of it does to attribution claims about a deployed model.

The authors report ELR collapse: learning rate and parameter norm govern loss dynamics primarily through their ratio, the effective learning rate. When ELR is matched across runs the loss trajectories collapse onto each other throughout training despite substantially different learning rates and parameter norms. Mean collapse errors across optimizers, architectures, datasets and model scales sit at a few times ten to the minus three, below the seed-to-seed variation measured in a representative configuration. Ablations point to normalization design and the timescale of learning-rate-versus-norm variation as the determinants.

Game development stresses coding agents unusually hard because program logic, visual and audio assets, interfaces, interaction and playability all have to work together in one executable artifact. Existing benchmarks score the final artifact or an isolated stage. Analysis of complete human-agent development trajectories identifies three stages spanning the lifecycle — initial generation, bug diagnosis and repair, and multi-turn optimization — and GameXpert-Bench evaluates all three rather than only the product.

Outcome-level evaluation tells you an agent run failed but not where the decisive error entered, leaving developers to inspect hundreds of steps by hand. Existing failure-attribution benchmarks use shorter traces. LongRCA Bench contributes 1,140 genuinely failed trajectories across five domains with no injected errors, independently scored human labels for the responsible role and the earliest decisive root-cause step, and a median trajectory length of 145 steps. The strongest baseline reaches 13.2 percent exact root-step accuracy.

Agent benchmarks routinely evaluate only final answers even when the agent runs on a stateful runtime, which under-specifies what is being measured: failures can occur in evidence acquisition, runtime routing, safety boundaries or repeated execution. ClawProBench argues the proper unit is a declared model-plus-runtime configuration, and instantiates that on OpenClaw with two tracks, a 102-scenario live profile with workspace tools and native routing tasks, and a frozen 68-scenario holdout with closed-world JSON contracts for stable ranking.

Conversational memory systems are evaluated with direct fact-seeking questions about prior dialogue. In a four-month deployment with 40 users, 1,872 sessions and seven memory conditions, Direct QA accuracy on existing benchmarks varied from 19.7 to 70.1 percent across conditions while satisfaction did not move. The authors argue the two track different capabilities: benchmarks measure elicited retrieval, recall when asked, whereas conversation requires natural integration, detecting relevance and weaving prior context into a response unprompted. MemUse is built from real user-cued memory moments to measure the latter.

Generative models are ranked by Fréchet and Kernel Inception Distance, but a first-two-moment summary misses distributional differences and a reported scalar gap is not a calibrated test against sampling variation. The concrete demonstration: on ImageNet, visually unrecognizable images optimized only to match the reference Inception mean and covariance obtain FID 24.7 against 58.6 for held-out real images. Both metrics are also symmetric under exchange of the two samples, so they cannot indicate direction of deviation. The paper offers detection, ranking and diagnosis procedures in place of the scalar.

Interpretable mixture-of-experts approaches such as Mixture of Decision Trees achieve transparency with homogeneous tree experts, which forces one inductive bias across every region of the feature space. This work adds decision trees, linear support vector machines and quadratic discriminant analysis as expert families under a shared probabilistic gate, calibrating the non-probabilistic experts to emit conditional class probabilities so parameter estimation stays inside the generalized expectation-maximization framework, with monotone ascent guarantees established for the heterogeneous gating updates.

Joint text-to-video-audio generation produces synchronized visual and acoustic content, but long sampling trajectories and heterogeneous multimodal computation make inference prohibitively expensive at scale. TurboT2VA accelerates a 19-billion-parameter joint model, addressing modality-imbalanced optimization with per-modality normalization, the difficulty of continuous-time consistency training at scale with a progressive curriculum, and the quality-versus-diversity trade-off that distillation usually worsens.

Full fine-tuning of EEG foundation models is impractical in resource-constrained clinical settings. The authors test whether updating only 9 percent of parameters through self-supervised adaptation suffices to align representations to a target task, across BIOT with a contrastive objective and CBraMod with masked reconstruction, on abnormality detection, event classification and seizure detection under both in-distribution and out-of-distribution conditions. Adaptation beats linear probing consistently, by up to 20 times AUCPR, and under a fixed compute budget peak performance needs only 20 to 50 percent of the available unlabeled data.

Text-based hallucination detection is well studied and spoken hallucination detection is not, particularly for low-resource languages. The benchmark contributes 12,013 news samples across English, Russian and Kazakh with controlled hallucinations at three types and three severity levels, each present as an original article and an aligned hallucinated counterpart in both text and audio, plus 290 natively collected fact-checked fake-news items in Russian and Kazakh translated and rendered through the same text-to-speech and speech-recognition pipeline. Evaluation compares fine-tuned multilingual encoders against zero-shot multimodal decoders on transcripts versus direct audio.

Almost everything known about post-training quantization comes from English benchmarks, which leaves open whether it transfers to morphologically complex low-resource languages. Evaluating Qwen-2.5-7B, LLaMA-3.1-8B and GPT-OSS-20B in full precision and in GPTQ-Int8, GPTQ-Q8 and GGUF-W8A16 across five Bangla understanding benchmarks, the three families diverge sharply: GPT-OSS loses up to 57.35 percent accuracy on reasoning-heavy tasks under GGUF-W8A16 while Qwen and LLaMA hold steady. Format choice, not just bit width, is doing the damage.

Agent tool invocation introduces file modification, information leakage and unauthorized action risks, and existing guardrails largely evaluate completed trajectories, leaving pre-execution step-level monitoring underexplored. StepGuard is a step-level guard model that both audits finished trajectories and checks individual tool actions before they run, trained on StepGen, an automatic data engine that generates safe and unsafe trajectories sharing the same context but differing in the decisive action, which is what makes the safety-utility balance learnable rather than reduced to blanket refusal.

The authors name a diagnostic illusion in retrieval-augmented generation: standard relevance proxies fail catastrophically on hard negatives, so measured evidence utilization is wrong before any allocation decision is made. They replace the proxies with an efficient causal leave-one-out probe that isolates generative reliance and formally calibrates the structural dilution of attention, then deploy the same probe in a deconfounded closed loop for allocating the context budget across retrieved candidates.

A retrieval-augmented system can return different answers after an index expansion even with model identifier, prompt, retrieval policy, evidence depth, rendering and generation controls all held fixed. Aggregate accuracy hides it when gains and losses cancel, and one-shot comparisons overstate update effects because generation is itself variable. The Snapshot Compatibility Audit estimates excess churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement, which separates the corpus effect from the sampling effect.

Generative world models are increasingly expected to replace physics engines, game engines and reinforcement-learning environments, but the remaining distance from generation to simulation has not been assessed systematically. This survey uses an external yardstick of eight simulator capabilities — asset construction, physics engine, interaction, controllability, stability, state feedback, diversity and evaluation metrics — and traces the main technical routes against each, which makes the gaps legible rather than aggregate.

Activation steering can change behavior without establishing that the effect is specific to the intended concept. SteerCheck is a preregistered attribution audit that matches off-target KL and separates mean, protected-tail, polarity, transfer and semantic claims. Exact replay of 960 Qwen3-14B interventions shows complementary limits in the two common controls: isotropic directions occupy a narrow near-orthogonal region, while sign-randomized same-construction directions often keep substantial target alignment, with effect strongly associated with signed cosine within that family at rho equal to 0.94.

When a context asserts two values for one fact a model commits to a cue — recency, repetition, position — but natural data rarely makes those disagree, so behavior alone cannot reveal which. The authors train 26-million-parameter transformers on a synthetic language where recency and rarity are exactly coextensive, then separate them with a minimal causal edit that inverts one cue while holding truth, token count and answer position fixed. All 75 runs reach accuracy at or above 0.999 including where the trivial heuristic fails, so no held-in evaluation distinguishes them, and under intervention the per-cell readout does not replicate.

Successful execution experience is typically discarded after a single run, so the next model rediscovers the same strategies and failure modes from scratch. EvoMap consolidates verifier-confirmed trajectories into structured genes, and LongWoF-Bench evaluates whether that transfers, with 778 machine-verifiable tasks spanning code generation, agent-environment synthesis, mathematical reasoning and rule following, including 252 tasks with verifier-confirmed Opus trajectories.

StarHarness evolves prompt and task framing, tool interfaces, skills, Model-Context-Protocol-backed providers, subagent structure and agent-loop configuration while leaving model weights untouched. It builds a compact evolution pool by stratifying tasks according to baseline failure behavior, separates proposer-visible search tasks from proposer-hidden selection tasks, and reserves held-out tasks to measure generalization rather than overfitting to the search set. Evaluation spans ITBench SRE, EnterpriseOps-Gym ITSM and AutomationBench Finance.

Multi-role, multi-stage agent workflows repeatedly transform upstream state into intermediate language artifacts — summaries, plans, tickets, memories, handoff notes — that downstream components then act on. For action-constraining state, topical retention is not enough: an artifact can still mention an unresolved condition while demoting it from a requirement that must be resolved before execution to information that may merely inform the next action. The paper studies this action-binding role as operational state preservation and shows safety blockers are among the things that weaken.

Web agents acting from rendered pixels avoid the fragility and token cost of reading HTML or accessibility trees, but training them needs large volumes of high-quality trajectories. Public datasets contain a few thousand trajectories from a fixed narrow set of sites, and even recent automated synthesis pipelines stay bound to predefined site lists or tutorial sources, so the number of distinct websites an agent ever sees barely grows. BrowserForge generates web episodes through parallel browser sandboxes to break that ceiling.

Deep research systems produce long-form cited reports by orchestrating multiple agents that search and synthesize from the web, and citations are the primary faithfulness signal, yet citation recall is poor. Improving it is hard because information passes through agents like a telephone game and both content and citations can be corrupted at any hop. The proposed evaluation pinpoints which agent introduced each error by testing each hop locally rather than scoring the report as a whole.

Evaluating retrieval-augmented generation requires assessing not just end-to-end correctness but how components interact and how errors propagate. This Bayesian framework jointly models retrieval success, abstention behavior and answer correctness, factorized along the pipeline's information flow, and separates whether the user got a correct answer from whether the generator behaved appropriately given the retrieval outcome. Applied across 27 configurations spanning three datasets, three retrievers and three generators, the conditional decomposition separates systems that look identical under marginal metrics.

In safety-critical operations, strong semantic scores do not guarantee operational reliability: a misread altitude, a dropped execution condition or a confused callsign can score well under standard F1 while carrying sharply asymmetric operational consequences. The authors instantiate consequence-aware evaluation in controller-pilot communication, where error tolerance is near zero, and use a controlled diagnostic setting to test how far semantic metrics diverge from operational risk.

Language models deployed as analysts are usually evaluated on what they can retrieve, not on whether retrieved information changes their judgment. Holding focal-firm information fixed and varying only unrelated context from 2,000 to 128,000 tokens, the authors find a risk disclosure's influence on investment judgments falls to the experimental noise floor while direct retrieval accuracy stays intact. The pattern replicates, which makes it a property of long-context integration rather than a retrieval failure.

Multi-teacher distillation normally trains the student to mimic teacher outputs on the union of the teachers' training sets, assuming that data is available. This work drops the assumption, first measuring how far one can get distilling from teachers fed different kinds of noise, then showing information already inside the teachers can be used to tailor the noise itself using an improved dead-leaves image model.

Preference alignment frequently causes catastrophic forgetting of pretrained general capability, and prior work frames that alignment tax as an optimization or architectural problem, leaving the properties of the preference data itself largely unexamined. BALIGN analyzes which characteristics of preference pairs drive the degradation, theoretically and empirically, and derives a balanced selection strategy that mitigates forgetting while preserving alignment efficacy.

Self-consistency samples diverse reasoning paths and picks the most consistent answer, and its main cost is token consumption. Existing efficiency fixes allocate paths by difficulty but sort problems into a few fixed levels, which does not capture the continuously varying nature of reasoning complexity. Flexible Self-Consistency estimates difficulty as a continuous signal and adjusts the number of sampled paths accordingly.

The paper views incremental grow-and-optimize training as progressive constraint relaxation: start from a low-dimensional submodel, then iteratively unlock nested random subspaces while freezing the orthogonal complement at initialization, re-optimizing after each expansion until the full architecture is reached. Framed that way, the question of why grown networks land in flatter regions becomes a question about the geometry of the constraint sequence rather than about stochastic optimization noise.

Clinical language models can produce recommendations that are factually plausible and physiologically unsafe. This framework couples a 7B clinical model with a hypergraph-neural-network physiological world model over an 847,000-node biomedical knowledge graph, scoring candidate responses against homeostatic constraints, multi-hop path plausibility and drug-interaction penalties, and using the resulting rankings to drive iterative on-policy ORPO updates. Evaluation runs on a 2,500-scenario Clinical Safety Benchmark built around physiological constraint violations.

Autonomous research systems can now execute long workflows, but automation alone does not keep the resulting process scientifically grounded. AutoResearch connects the two halves: idea generation continuously integrates emerging research signals with accumulated domain knowledge, identifies transferable mechanistic insights, and uses multi-model generation with cross-review to produce testable plans; idea execution decomposes those plans into experiments, iteratively implements and diagnoses them, and applies independent evidence-based review before a result is accepted.

Spoken queries pass through automatic speech recognition before any retrieval, so ASR errors enter as a fixed upstream constraint. Testing whether entity-graph linking and iterative reformulation absorb or amplify those errors, across four synthesized English accents and four RAG configurations on HotpotQA, 2WikiMultiHopQA and MuSiQue against a clean-text oracle, the answer is amplify: the structurally richer configurations retain higher absolute F1 under ASR input, but the gap from clean text to the highest-word-error-rate accent is 36 to 67 percent larger under their combination than under naive dense retrieval.

Video multimodal models supporting language-guided segmentation show jitter, drift and identity switches, especially with partially hidden targets or nearby similar objects, and one likely reason is that training carries no explicit spatial prior to hold identity and shape stable over time. PhysMLLMs injects physics-inspired spatial continuity priors at training time, aligning the student's global visual representation with a frozen teacher through global representation prior alignment to encourage object-centered stability.

The de facto reference for morphological transforms in Python, scipy.ndimage, is CPU-only and single-array, so using it inside a GPU training loop costs a device-to-host round trip, and PyTorch vision libraries cover only a narrow subset, usually two spatial dimensions and flat structuring elements. TorchMorph exposes 22 public operators spanning binary and greyscale morphology, exact and approximate distance transforms, and entropy-regularised optimal transport, implemented as fused CUDA kernels operating directly on batched channel-spatial CUDA tensors with up to eight spatial dimensions.

Industrial technical reports hold high-value maintenance, troubleshooting and product-engineering knowledge, but their heterogeneous structure of dense prose, specifications and tables makes them hard to index and reason over with standard pipelines, and no public instruction-tuning or benchmark data is built from them. Using 906 public Panasonic documents across 7,525 pages, the authors apply layout-aware extraction, build a semantic retrieval index, and synthesize multiple-choice questions grounded in retrieved evidence under five query-document relationships including irrelevant retrieval and single- versus multi-document support.

An analysis of recent promotions in the Russian general officer corps and what the pattern suggests about which operational approaches are being institutionally rewarded. Not an AI item, but it comes from a configured defense source inside the window and is included for completeness of the coverage matrix.

Agents are moving past conventional retrieval-augmented generation toward direct interaction with a full corpus, but reachable evidence can stay unusable under a finite interaction budget: required evidence may never surface, a surfaced supporting document may go unopened, or an opened document may not expose its decisive fragment. The authors call this progressive silent loss evidence blindness and quantify it through stage-wise evidence realization, arguing that raw interaction adds little reusable corpus structure across queries.

Instead of an open-ended search loop, Crase queries a search engine once for seed papers, expands along their 1.5-hop citation neighborhood, prunes citation edges whose claims lack entailment support, and ranks what remains with a recency-aware random walk. The candidate set, the reason each paper is kept, and the stopping condition are all explicit and fixed before inference. On LitSearch and one further benchmark over a 500,000-paper arXiv corpus it outperforms deep research agents built on proprietary models by up to three times recall at 50, at roughly a third of the cost.

Uncertainty quantification is widely used for closed-book hallucination detection, and prior work proposes combining UQ signals in learned ensembles, but robustness has not been examined carefully. This study trains a classifier over heterogeneous UQ scorer outputs on a small domain-specific labeled set and applies it out of sample with no retrieval, tools or reference, isolating when the ensemble genuinely helps and when the gain was an artifact of in-domain fitting.

Neural emulators trained on solver-generated data promise large speedups over numerical partial-differential-equation solvers but are usually framed as opaque alternatives to the very methods producing their training signal. This thesis argues the two paradigms are closer than that: neural architectures mirror classical discretizations, their errors submit to the same spectral analysis, and insight flows in both directions rather than only from solver to emulator.

Individual-fairness audits compare predictions among similar individuals in feature space, which makes the audit's verdict a function of nearest-neighbor relationships. Small perturbations can therefore reshuffle local neighborhoods and change the fairness assessment even when the model's predictions are unchanged. The paper develops a geometric analysis of audit robustness under bounded perturbations and establishes sufficient conditions for stability.

Model misspecification degrades Bayesian optimization and active learning, and fully Bayesian treatments that would fix it usually require expensive Markov-chain Monte Carlo sampling. KENDO substitutes an ensemble of Gaussian processes over different kernels with adaptive Bayesian weighting, paired with disagreement-aware acquisition, and instantiates the same self-correcting mechanism as KENDO-BO for optimization and KENDO-AL for active learning with task-specific acquisition.

Continuous-control policies are commonly optimized as unbounded Gaussians and then mapped into bounded actions, and the paper shows the choice of where entropy is measured changes the geometry the policy converges to. In an 80-muscle MyoLeg task a clipped Gaussian executes 89.07 percent of actions within 5 percent of a bound; a same-state decomposition rules out variance as the sole cause, since zeroing variance still leaves 83.83 percent of actions near a bound and 82.12 percent of state-conditioned means fall outside the executable interval. Replacing clipping with a tanh map does not remove the high-variance regime.

Multi-task vehicle-routing solvers are limited on two fronts. In training, reinforcement learning suffers reward-scale disparities and shrinking advantage signals as the policy improves, while preference optimization stagnates once sampled tours become near-identical, leaving both with weak late-stage supervision. In architecture, fully shared encoders entangle constraint-dependent representations across heterogeneous variants and limit generalization. The paper addresses both with locally augmented preference construction and representation disentanglement.

Conversational AI systems change continuously through model releases, feature updates, safety interventions and access-policy shifts, but user perception is usually studied as a static snapshot. Combining sentiment classification with thematic concept analysis over long-run Reddit discussion, the authors find perceptions are dynamic and intervention-sensitive with distinct provider profiles: Anthropic shows the clearest positive release profile driven by Claude Code and product-model fit, OpenAI shows backlash-and-recovery dynamics around GPT-5 and GPT-5.1, Grok-3 is shaped by provider identity, and DeepSeek-R1 pairs engineering praise with concerns about censorship, access and reliability.

Multimodal models inherit backdoor risk from their construction pipelines, with triggers potentially residing in images, text or both. Model-level removal methods designed for conventional classifiers transfer poorly, and multimodal-specific defenses mostly operate at inference time, filtering suspicious inputs while leaving the backdoor in the weights. RACER is a model-level repair framework built on a region-aware consistency observation about where backdoor behavior concentrates, aimed at eliminating the latent backdoor at its source.

Stability AI has closed its Series B under new leadership, bringing total funding to 232 million dollars, with the round adding Electronic Arts, Sony Music Group, Universal Music Group and Warner Music Group among others. TechCrunch reported the new capital at 76 million dollars. The investor list is the story: each of these companies already has a bilateral partnership with Stability, and the round converts a set of licensing and co-development relationships into equity alignment. It also lands a week after Stability shipped a digital-audio-workstation plugin and a more advanced generation experience for Stable Audio 3.0, the fully licensed-data audio family it released in May.

A systematic sample of daily top-five Hacker News stories, run in February 2026 and repeated in June, finds AI-related or AI-generated content rose from about 40 percent to roughly 60 percent in the first half of June, settling near 50 percent by month end. The author used Pangram to flag likely machine-written submissions and manually reviewed every flagged story, reporting a few false negatives and no obvious false positives. The June pass distinguishes pure AI navel-gazing, meaning vendor announcements and opinion pieces about the technology, from stories that lean heavily on AI but carry broader consequences, and excludes items only tangentially related such as memory price reports.

Ben Lorica's newsletter walks through what high-bandwidth flash and tiered memory hierarchies mean for inference cost structures, at a moment when memory pricing has been the dominant infrastructure variable for most of the year. The piece pairs usefully with today's two accelerator disclosures, both of which are explicitly organized around minimizing data movement rather than adding arithmetic.

Sam Charrington's conversation with Max Welling covers the case that physical priors and physics-derived formalisms, rather than further scaling of general-purpose architectures, are the more likely source of the next architectural step change. Welling has been making a version of this argument for years; the current interest is that it lands while several groups are shipping physics-informed priors into otherwise conventional systems, including today's PhysMLLMs paper.

GitHub's engineering blog published a long practitioner walkthrough of pre-production language model evaluation: building task-representative eval sets, choosing between automated scoring and human review, handling non-determinism across runs, and deciding what gates a release. It is aimed at teams shipping model-backed features rather than at researchers, and it is a reasonable counterweight to the several papers in today's digest showing that benchmark scores and deployed behavior diverge.

An interview with OpenAI head of product Thibault Sottiaux, framed around the argument that consumer and enterprise readiness rather than model capability is now the binding constraint on agentic product adoption. It arrives the same week the company disclosed its first custom-silicon benchmarks and lost another infrastructure executive, so the product-readiness claim is being made against a noisy operational backdrop.

Anthropic's Cowork agent now retains context supplied conversationally rather than losing it between sessions, closing a gap that made repeated instructions necessary for recurring workflows. It is a small product change but it sits on the same axis as this week's memory research: the useful capability is not recall on demand but integration of prior context into work without being asked.

Accel-backed Keenable is building a web index whose retrieval surface is shaped for agent consumption rather than human browsing, on the premise that agents pay a large token and latency tax reading pages designed for eyes. It is a thin report, but the category is worth tracking alongside Perplexity's Search-as-Code interface and the BrowserForge result on how little of the web any web agent has actually seen during training.

Cohere released an IDC InfoBrief reporting rising urgency among global organizations seeking more control over their AI deployments, which is the market position the company has been building toward all year alongside its EU Code of Practice signature in July and its Command A sovereign-agent line. The brief itself is vendor-commissioned market research rather than independent measurement, so the interesting content is the framing: sovereignty is being sold as a control and jurisdiction requirement rather than a performance one.

Ringg has raised from Peak XV to extend its voice AI stack beyond telephony into other conversational surfaces. Indian-market voice deployment is a useful signal for multilingual speech systems generally, since the language coverage requirements are considerably harder than a single-locale product needs.

Gamma has acquired Lica, an Accel-backed design startup, folding its team and product into Gamma's generative presentation and document tooling. Small consolidation in the generative-design tier, which has been steadily absorbing point solutions all year.