Wolf Digest — Tuesday, August 18, 2026

The Air Force evaluated prototypes of a portable Command and Control Enclave at Springfield Air National Guard Base in Ohio, the initial assessment phase of a previously unrevealed capability under the Collaborative Combat Aircraft program. The C2E is intended to let airmen control CCA drones from the ground using a transportable system rather than fixed infrastructure, which changes the basing and sortie-generation math for autonomous wingmen. The service says it is now closer to selecting a prime contractor for the effort.

Generalist vision-language-action models do not transfer cleanly to humanoid whole-body loco-manipulation, because coordinating locomotion, waist posture, and dual-arm manipulation exceeds what single-stage VLA architectures handle well, and offline behavior cloning leaves policies suboptimal at deployment. HAF splits the problem into a VLA component and a student controller, then refines online in a spectral latent space rather than tuning the large backbone directly — which keeps real-robot exploration cheap enough to run and bounded enough to be safe. The two-part structure is the general pattern this subfield has converged on: freeze the expensive generalist, learn the cheap adapter.

Wiz researchers show a chain in which an AI-generated Copilot 'autofix' suggestion, once merged, granted enough continuous-integration privilege to compromise a Snowflake Jira instance. The interesting property is that no individual step required a traditional exploit: the vulnerability lives in the trust boundary between an agent that can open pull requests and a pipeline that runs merged code with elevated credentials. It is the clearest recent illustration that agent write-access to a repository is effectively write-access to everything the pipeline can touch.

The now-standard recipe for long-horizon manipulation freezes a VLA, puts an LLM agent in charge of planning, uses analytic primitives for free-space motion, and invokes the VLA only for contact-rich segments. BATON identifies two ways this breaks at long horizons: whole-task exploration cost is multiplicative in stages, so a K-stage task needs roughly T-to-the-K episodes and a failure does not reveal which stage caused it; and the representation carries exit conditions but no entry conditions, so one subtask silently constrains the next. The fix is per-subtask exploration plus a memory that records transitions rather than outcomes.

Reporters hid an AirTag in a shipment of rare books and tracked it to an Amazon facility where physical volumes are destructively scanned for AI training data. The underlying economics are not disputed: models have largely exhausted openly available web text, so out-of-print and never-digitized print corpora are among the few remaining sources of high-quality, non-synthetic, non-contaminated text at scale. The dispute is over destructive digitization of scarce physical objects and over whether the resulting corpus is licensed at all.

Listed pricing for GPT-5.6 Sol was cut by half, a move visible first through routing-broker price pages rather than a lab announcement. Halving the price of a top-tier reasoning model compresses the cost-per-task gap against cheaper models and, for agentic workloads where a single task can consume tens of thousands of reasoning tokens, changes which model is economically rational for long-horizon work. It also lands the same week a payments company moved to acquire the largest neutral routing layer, which is where price changes like this become immediately visible to buyers.

When an LLM drives a robot, the environment state it reads back is an untrusted input channel, and this work formalizes attacks that inject adversarial semantics into that channel rather than into the user prompt. The threat model is more realistic than prompt injection for physical deployments: an attacker who can alter an object, a label, or a sensor reading in the workspace does not need access to the agent's text interface at all. The lineage runs through affordance-grounded planners and code-as-policies systems, all of which trust perceived state implicitly.

Anduril has delivered redesigned Soldier Borne Mission Command headsets to the Army for testing and is preparing for the next prototyping phase, according to a company executive. SBMC develops soldier-worn heads-up displays combining augmented reality, night vision, on-device AI, and command-and-control links to other systems including drones. The program is the successor effort to the Army's troubled Integrated Visual Augmentation System.

The authors demonstrate that individually weak and seemingly irrelevant prompt cues — paraphrases, typos, incidental stylistic choices — can be systematically combined to exert strong control over model behavior, an effect they call model hypnosis. It appears across model families and scales including frontier reasoning models, and hypnotic prompts transfer between models, which implies the mechanism is a shared property of the training distribution rather than an artifact of one alignment recipe. The transferability is the part with teeth for both safety and interpretability: a control channel made of inconspicuous textual choices is hard to filter and harder to attribute.

An investigation describes the creation of a think tank with no apparent staff or research history whose output appears designed to be ingested and cited by AI chatbots. The mechanism is retrieval poisoning at the reputation layer rather than the document layer: rather than injecting adversarial text into a corpus, the approach manufactures a source that assistants' retrieval and citation heuristics treat as legitimate. It is the practical counterpart to a growing academic literature on retrieval-augmented generation poisoning, two examples of which appear elsewhere in today's set.

Nvidia is putting 1.5 billion dollars into the SoftBank-affiliated data-center developer building an OpenAI facility, an investment that effectively guarantees Nvidia silicon powers the site. The pattern — chip vendor taking equity in the developer of the buildings that will house its chips — is now a recurring structure in AI infrastructure finance, and it blurs the line between supplier and customer in a way that makes reported demand harder to read.

Existing embodied agent harnesses are largely open-loop: they follow fixed skills during a rollout and reflect only after the episode completes, which cannot govern execution while it is happening because physical interaction changes state faster than a large agentic model can deliberate. Zetta proposes a closed-loop harness that separates fast execution-time control from slower agentic reflection, so the agent can intervene mid-episode without running the large model at control frequency. It is the embodied counterpart to the harness-optimization work appearing on the language side today.

This work instantiates two normative frameworks as differentiable reward signals for an autonomous-driving policy in CARLA: a utilitarian objective minimizing total casualties, and a Kantian objective treating course maintenance as a categorical imperative. The policy is trained with proximal policy optimization against a Bradley-Terry reward model fit to human preferences, which makes the ethical framework an explicit, swappable training artifact rather than an emergent property of the driving objective. Whether encoding trolley-problem framings as reward is the right abstraction for real driving is a fair objection; making the encoding legible is nonetheless progress over leaving it implicit.

Letting a vision-language-action model judge the reliability of its own action generation, without expert annotation, has mostly been attempted from output statistics. This work instead reads internal visual-modality entropy, observing that it varies consistently with action quality across architectures, and builds a self-evaluation signal from that. Internal-signal confidence estimation is attractive for robotics precisely because the alternative — discovering the failure by executing it — is expensive.

NebulaVLA decouples high-level semantic reasoning from low-level action control into asynchronous frequency bands, which is how the efficiency-versus-performance trade-off in deployed VLAs is usually broken. To handle cross-embodiment transfer it introduces GESTURE-7, a unified language-grounded action representation shared across heterogeneous robots, plus a guide-action mechanism intended to smooth execution across the frequency boundary.

Conventional on-orbit planners rely on predefined maps and fixed environmental assumptions, which do not survive contact with dynamic orbital scenarios and limited onboard observation. Orbit-Planner learns action-conditioned spacecraft dynamics and rolls out future states in latent space, adding a physics probe to keep the learned dynamics consistent with orbital mechanics rather than merely plausible. Learned world models for spacecraft are a natural fit for the regime where onboard compute is scarce and ground-loop latency is fatal.

Existing surveys of embodied-agent security organize threats by mechanism — jailbreaks, prompt injection, backdoors, poisoning, adversarial examples — which does not tell a system designer where an adversary first enters the control loop. This survey reorganizes the field around trust boundaries, tracing how a compromise at each interface propagates from digital input to physical behavior, and maps existing defenses and evaluations onto those boundaries.

Groq raised 350 million dollars at a 3.5 billion dollar valuation to fund its shift from selling custom inference silicon to operating a neocloud, including expansion of an Nvidia-powered data-center footprint. A custom-accelerator company buying general-purpose GPUs to serve capacity is a notable admission about where inference demand actually is: the customers want tokens, not architectures, and the fastest path to revenue is capacity rather than differentiation.

Judging a world-model rollout requires deciding whether physics, causality, and world state evolve correctly, which existing benchmarks reduce to brute-force metrics that leave no examinable reasoning chain. HarnessEval-W interprets each evaluation case, decomposes the judgment into measurable subproblems, and spawns specialized sub-agents with tailored context and diagnostic tools for each. The trade is auditability against cost and against the reliability of the judging agents themselves, which the paper has to argue for rather than assume.

Pixel-space diffusion has been studied mostly at small scale or in class-conditional settings, leaving no practical recipe that rivals latent-space models. The authors first confirm that direct large-scale pre-training in pixel space converges substantially more slowly than in latent space, then propose a latent-to-pixel strategy: acquire generative priors efficiently in latent space, then transition to pixel space during post-training. They systematically ablate the transition choices — weight initialization, data composition, prediction target, decoder architecture, and noise schedule — which is the part practitioners will reuse.

A detailed commentary on recently published monitorability evaluations argues that reading a model's chain of thought remains among the most promising detection methods for undesirable behavior, but that current evaluations do not establish what they are taken to establish. The core objection is that monitorability results are conditional on the model having no incentive or ability to obfuscate, and that existing eval designs do not separate 'the reasoning was legible' from 'the model had no reason to hide anything'.

CHIVE is an agentic pipeline that finds unexpected model behaviors in the wild and investigates them with counterfactual prompt edits, producing thousands of explanations paired with supporting counterfactual evidence. Evaluating explanations by counterfactual simulatability — does the explanation help predict behavior on related inputs? — the authors report the uncomfortable finding that common interpretability techniques do not reliably improve an agent's ability to make those predictions. That is a direct challenge to the working assumption that mechanistic explanations are automatically actionable.

Agentic image-generation pipelines add retrieval and iterative reasoning on top of text-to-image models but run fixed one-size-fits-all topologies, so simple prompts get pushed through computationally heavy workflows. GenRouter first defines GenCanvas, a standardization of diverse agentic pipelines into a set of foundational primitives and executable templates, then routes each request across that unified space according to difficulty. Routing by query difficulty is the same economic idea driving model routing at the API layer, applied one level down inside the workflow.

A controlled study varying one generalization factor at a time — in-domain shift, cross-domain transfer, multi-teacher — finds that on-policy distillation transfers a teacher's reasoning behavior rather than its answers to particular problems. Training difficulty barely matters, and problems the teacher never solves are still useful training signal. The strongest determinant is the origin relationship: same-origin teacher-student pairs transfer across languages, reasoning horizons, and even other domains, while cross-origin pairs mostly fit the trained distribution. That is a practical constraint on distillation shopping, and it appears alongside three other on-policy-distillation papers in today's arXiv batch.

Automated checking pipelines increasingly put one model in the checker role and another (or the same one) in the fixer role. Holding the present task byte-identical, the authors measure false alarms on human-verified-correct ProcessBench traces and find that a completed audit-then-repair episode already in context lowers false-alarm rates in all fifteen model-by-wording combinations tested, by 2.8 to 11.5 percentage points against a length-matched non-audit control. The implication for anyone running verifier-in-the-loop systems is that verification strictness drifts with conversational history, so a verifier's calibration is not a fixed property of the model.

Shield AI describes the authorization process behind a Frontex maritime surveillance mission flown from the Italian Coast Guard patrol vessel Dattilo, in which the Italian Civil Aviation Authority evaluated not just the V-BAT airframe but the specific configuration, the crews, the operating procedures, and the mission plan. The write-up is a useful counterweight to autonomy demos: for uncrewed systems operating in civil airspace, the binding constraint is regulatory evidence about the whole operating system, not the autonomy stack in isolation.

Third-party providers hosting open-weight models are now a large part of the serving ecosystem, and buyers have no direct way to verify that the endpoint they pay for is running the model, precision, and configuration advertised. Ventor-QTest formalizes hosted model routing as a stochastic process and proposes a composite black-box audit requiring no probability information from the target API: repeated requests against frozen constrained contexts let the auditor reconstruct enough of the output distribution to detect substitution and quantization. It pairs naturally with the endpoint-accuracy indices independent evaluators have begun publishing per provider.

Frontier models solve hard tasks once the problem, tools, and success criteria are specified, but consequential real-world problems rarely arrive in executable or verifiable form. Apodex Discovery surveyed 561 industries across 16 sectors, assembled 423 high-value real-world problems, and selected 20 to build into stateful, verifiable investigation environments driven by what the authors call a heavy-duty solver: foundation model, harness, tools, and control policies. The problem-scouting methodology is arguably the more transferable contribution, since the shortage in this area is well-specified real problems rather than solver capacity.

Regulatory compliance monitoring is increasingly implemented as an automated control that checks model outputs against written rules spanning data protection, healthcare, financial regulation, and platform policy. This audit of activation probes and guard models shows the verdicts frequently do not depend on the stated rule at all: deleting, permuting, or substituting the rule text leaves detector output largely unchanged, a failure the authors call rule blindness. If a compliance detector fires on scenario surface features rather than the regulation it cites, its audit value is close to zero.

Using circuit analyses across 46 tasks spanning language, formal reasoning, social reasoning, and physical reasoning, the authors find that tasks recruiting the same network in human brains recruit overlapping neurons in language models, while tasks recruiting different human networks recruit distinct neurons. The claim is that functional modularity mirroring the human brain emerges from a very different optimization process, which bears on whether such specialization is a convergent property of intelligent systems or an evolutionary accident. The obvious caveat is that shared training data about human cognition is itself a confound.

Scalable post-training for biological reasoning has depended on costly hand-curated reasoning traces. PertMind instead treats cellular perturbation atlases as reinforcement-learning environments, using measured gene responses as computable rewards at the gene, pathway, and format levels on top of a trusted-trajectory supervised initialization. Trained only on forward perturbation-response prediction, it transfers without task-specific post-training to reverse perturbation identification, double-perturbation reasoning, phenotypic-screen prioritization, and biological-process interpretation — which is the claim worth checking, since zero-shot transfer across those four is a strong statement about what the reward taught.

Can a model recover a paper's true research idea given only that paper's pre-publication bibliography? Reconstruction withholds the seed paper and all contemporaneous or later literature, enforcing a temporal citation cutoff, anonymous reference identifiers, and frozen per-paper bibliographies to block prompt-time leakage. Across six domains and 643 papers, seven frontier models achieve match rates of only about three to fifteen percent, which is a useful cold-water number for claims about automated idea generation.

Most reasoning benchmarks give each task its own budget, which hides the resource-rationality question: given one shared budget across six problems, does a model allocate it well? R³-Bench calibrates each suite against the same model's demonstrated single-problem competence, and across 72 model-by-condition cells the offline oracle matches or exceeds the contest mean in every cell and strictly exceeds it in 71. Under moderate tool-free pressure, even equal allocation beats the model's own choices for four of six models — that is, the allocation policy is worse than not having one.

Foundation GUI agents are held back by scarce and biased training data, ambiguous prompts, and unreliable execution, and routine workflows depend on user-specific tools and tacit conventions that unstated instructions cannot recover. UI-Mate pairs a closed-loop data engine — automated task generation, environment construction, rollout, filtering, capability balancing, supervised fine-tuning, and online reinforcement learning across massively parallel environments — with a mechanism that turns multimodal demonstrations into flexible subtask-level workflows the agent can follow selectively.

Clinical decision support is moving toward committees of language-model agents deliberating in a shared workspace. Across seven cohorts on six public datasets spanning text, imaging, and tabular ICU records, individual agents resist benchmark-gaming cues in isolation with flip rates of five to sixteen percent — but a socially plausible shortcut introduced into the shared workspace spreads through the committee. The finding argues that multi-agent deliberation can amplify rather than filter a shortcut when the shortcut is socially credible.

Answer-only reinforcement learning trains models on fully specified problems, but real queries routinely omit a premise needed for a unique answer, and refusal is usually the wrong response. ACA-RL builds missing-premise training instances from well-posed problems using a reasoning-graph-guided pipeline with localized gap annotations, then trains with a structured reward over five observable behaviors: asking for the missing premise, conditioning the answer on the unknown, abstaining, and two failure modes. The accompanying Missing-Premise Benchmark contributes 274 human-verified instances across mathematical, logical, and real-world settings.

Global-workspace framings of language models invite an admission story: some gate decides what enters reportable form. Testing that on open-weight models with Jacobian lenses across a five-arm benchmark sharing an identical context, the authors find no gate where the story predicts one. Task demand raises a concept's lens visibility — roughly five percentile-rank points on the primary checkpoint, positive on all four measured — while a single shared linear map decodes the variable from every arm including the control at six to nine times its selection-corrected floor. The reading is that verbalizability is a matter of degree along a shared direction, not an admitted-or-not binary.

Video world models approximate a distribution over physical outcomes, but existing benchmarks score individual generations or compare distributions coarsely, leaving aleatoric uncertainty untested. CaliBench scores outcomes in physically interpretable discrete spaces — a bin index, a die face, a suit, a color — whose reference distributions are known in closed form, including binomial Galton boards, Bernoulli forks, uniform dice and cards, and a skewed European-roulette color. That permits an exact calibration test, and the authors decompose results into scorability and calibration, two axes a single accuracy number conflates.

Repository-scale coding requires holding tests, imports, configuration, and migration rules consistent inside a bounded context window. The authors model this as reconstructing a coupled-fact graph in which each edit needs a fact drawn from recent context or parametric memory, with facts covered by neither forming coherence debt. Supplying and withholding each channel across seven models and five harnesses, they confirm no model completes a task on an unseen API with both channels empty — and, more usefully, quantify how the two channels substitute for each other.

Group-relative methods such as GRPO cut gradient variance by sampling many rollouts per prompt, but give only sequence-level credit and stall on straggler rollouts, which reduces throughput and increases off-policyness. Learned value functions solve both problems in theory — token-level advantages without large groups — but the infrastructure cost and the practical success of critic-free methods have kept them out of production pipelines. This work proposes privileged value functions, which inject task-relevant token-level signal available at training time but not at inference, alongside a complementary technique aimed at the straggler problem.

Mint-Agent is built around two demands financial agents face that general agents mostly do not: precise operations over grounded evidence, and long-horizon research whose conclusions remain auditable afterward. The system pairs a data engine that constructs atomic-capability and long-horizon tasks from real financial sources with MintHarness, which maintains evidence trails across extended trajectories, and a training recipe combining supervised fine-tuning, critical-step on-policy distillation, and reinforcement learning with verifiable rewards, unified by model merging of separate reasoning and execution experts.

Nvidia lays out how it is extending the supply-chain discipline it applied to advanced packaging, memory, and networking to the physical inputs of AI data centers — land, power, and shell capacity, which it abbreviates as LPS. The framing that 'compute is revenue' is a vendor line, but the underlying claim is checkable and increasingly the binding constraint: interconnect queues and substation lead times now gate deployment more often than accelerator supply does.

Palmyra x6 post-trains a mixture-of-experts base with anchored supervised fine-tuning on a deliberately tiny corpus: 626 verified synthetic tool-use trajectories, a single epoch, a low learning rate, a KL anchor to the frozen base, and a Muon-plus-Adam hybrid optimizer. It reports the cohort's highest BFCL Core score at 0.785 and the highest six-benchmark mean. The interesting claim is the data efficiency — that agentic tool-use competence can be elicited rather than taught, if the base model already has it and the fine-tune is anchored tightly enough not to damage what is there.

Memory-based sequence models compress context into a compact state, but nearly all of them expose a static memory throughout the sequence, so early tokens face no compression pressure, occupy too many degrees of freedom, and crowd out later context. Proteus instead grows effective memory capacity as the context grows: an early bottleneck forces aggressive compression of history, and capacity unlocked later reduces interference between stored and arriving content. It is a simple scheduling idea applied to a component the linear-attention and state-space literature has mostly treated as fixed.

Multi-preference alignment is usually framed as scalarization — combine reward dimensions, then optimize — which leaves the temporal question unspecified: when should each preference dimension enter optimization at all? STAGE is a stability-guided active-set controller that starts from a small set of objectives, retains admitted ones, and expands when reward-deviation gates indicate low recent deviation or a patience budget expires, with a probing phase estimating a hard-to-easy ordering. Across fifteen training preferences and sixteen held-out benchmark columns it beats simultaneous scalarization and shared-budget baselines on average.

Unlike classical bandit agents, LLM agents see textual labels attached to actions, and those labels carry pretraining associations with no formal counterpart in the task structure. The semantic bandit setting isolates that effect and shows semantically informative labels reduce exploration in favor of exploitation — which helps when the semantics align with the true reward and hurts sharply when they do not. It is a clean diagnosis of why LLM agents look decisive on familiar-sounding action spaces and lost on renamed ones.

The authors organize risks from expanding agentic cognition into a three-level framework by cognitive scope — physical cognition, social cognition, and self-referential cognition — and map each level to the human capacity it stresses: agency, autonomy, and control. It is a taxonomy paper rather than an empirical one, but the scoping axis is more useful than the usual capability-versus-alignment split because it points at where a mitigation has to sit.

Existing evaluations of agentic 3D world construction use idealized simple queries, which cannot distinguish intent understanding from tool use from multimodal feedback reasoning. VibeWorlding contributes VWE-BENCH: 2,616 curated 3D assets, 323 human-annotated seed worlds, and 6,828 reverse-synthesized multimodal user queries, split into verified queries with ground truth and unverified queries designed to probe open-ended behavior.

Output-side detectors for retrieval poisoning rely on perplexity and consistency checks, but the authors show deliberate attacks often induce false confidence — poisoned outputs exhibit lower perplexity than benign ones — which inverts the signal these detectors depend on. Their alternative reads document-level attention collapse instead, detecting the concentration pattern that a successful injected document produces regardless of how fluent the resulting output is.

As AI data centers retire still-functional accelerators, large volumes enter secondary markets. The authors physically built a 128-GPU cluster entirely from second-hand components and ran it for a year serving modern LLM inference, reporting the capital cost and the conditions under which such repurposing is economically viable and environmentally defensible. The result is a useful floor on what inference capacity costs when you are willing to accept older memory bandwidth and no vendor support.

Part-aware 3D generation underpins controllable modeling, editing, and articulation, but existing methods do not scale: as part count grows, token length and memory make detailed geometry prohibitively expensive. MegaParts is an autoregressive framework that reaches 300 parts per object through a token-efficient representation, which is the regime where generated assets start being usable in articulated simulation rather than only as static meshes.

This agent factors long-horizon household tasks into two phases: a vision-language model and exploration harness acquire goal-relevant predicates and instance bindings from egocentric observation to produce a symbolic initial state, then a PDDL transition model restricts decoding to tokens extending applicable actions while Monte Carlo tree search evaluates continuations with a domain-independent heuristic. Plans are executable by construction under the transition model, with real-world transfer conditioned on the grounding being right — which relocates rather than removes the hard part.

Group-relative reinforcement learning is now the default for post-training reasoners, but when multiple reward objectives are involved, the standard practice of scalarizing the reward vector before group-wise standardization creates two problems: rollouts with distinct reward profiles receive identical advantages, and all objectives are optimized at fixed relative weights regardless of how saturated each already is. The proposed reweighting concentrates gradient on objectives that still have headroom, which is the multi-reward analogue of curriculum ordering.

Artificial Analysis added an independent evaluation of Qwen3.8 27B to its Intelligence Index, which currently aggregates nine evaluations including GDPval-AA v2, a banking agentic tool-use suite, Terminal-Bench, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, an omniscience and hallucination measure, and a long-context reasoning test. Independent per-endpoint evaluation matters increasingly for open-weight models, where the same weights served by different providers can differ measurably in accuracy.

Finite noisy data routinely admit expressions that fit equally well but imply very different behavior outside the observed regime, so numerical fit and syntactic complexity are insufficient measures of scientific credibility. AutoSR searches over persistent scientific investigations — accumulated research states carrying evidence, assumptions, and prior attempts — rather than over expressions in isolation, which is an attempt to encode the part of symbolic regression that a scientist supplies by hand.

Multimodal retrieval-augmented pipelines for visually rich long documents have drifted toward multi-agent systems, but with fixed workflows that cannot scale computation at test time, so hard queries get the same evidence budget as easy ones. D2-ScaleAgent adds a verifier-driven dynamic routing loop around a continuously updated evidence bank serving as working memory, expanding retrieval breadth or reasoning depth separately depending on which the verifier judges insufficient.

Enterprise question answering is usually framed as retrieve-then-generate over internal documents, but routine corporate records are work by-products in which the organizational relations a question depends on remain implicit and scattered. Existing benchmarks materialize a predefined answer path and therefore test composition of stated facts; ENTLORE targets latent organizational reasoning — recovering a relation that is nowhere written down — using a graph-grounded construction.

Conventional AutoML searches inside a space of preprocessing operators, learners, and hyperparameters fixed in advance, so it can select and tune known components but cannot produce structure outside the space. LACE maintains a population of scikit-learn-compatible Python classes and uses a language model as the evolutionary variation operator, making the search space the set of executable programs rather than a configuration grid.

Google Research describes models that estimate cardiometabolic risk markers from smartphone imagery, positioning the approach as a refinement on body mass index, which is a crude scalar that conflates body composition. The upside is a low-friction screening signal from a device nearly everyone carries; the risks are the usual ones for imaging-based health inference — demographic generalization, calibration drift across camera hardware, and the consequences of a screening tool being used as a diagnostic.

The essay argues the standard analogy between open language models and foundational open-source software projects such as Linux paints a narrower path than people assume. Linux became self-sustaining once it was clearly the best tool for the job and contribution was cheap; open language models require recurring capital for each training run, so the self-fulfilling dynamic that carried Linux does not transfer without a durable funding mechanism attached to the training loop rather than to the artifact.

Translating natural language into the Planning Domain Definition Language lets symbolic planners produce verifiable plans and sidesteps the long-horizon unreliability of direct LLM planning, but existing methods use rigid pipelines, partial domain definitions, or human feedback, and are evaluated without standardized automated verification. PDDLCoder contributes an agentic generation loop plus a benchmark with automated verification, which is the part the subfield actually lacked.

Most work on AI scientists studies the autonomous capability of the agent, which the authors argue misses the social structure of scientific work. They propose that the human-agent pair, not the agent, is the right unit of analysis, and support the case with literature review, empirical analysis, and recent incidents where the failure was in the collaboration rather than in the model.

Token-level on-policy distillation gives fragmented corrections along an erroneous student trajectory and cannot unfold a complete repair path. SOPD supervises at the step level over complete student-generated trajectories, and the authors show it reduces to supervised fine-tuning or approximates on-policy distillation at the two limits of step length. Because teacher responses are conditioned on the student's own trajectory, the supervision stays on-policy while recovering long-horizon corrective structure.

Benchmarks for multi-reference image generation are organized around predefined task labels such as subject composition, which fragments coverage and gives no diagnostic signal in what is fundamentally a combinatorial setting. TRACE-Bench takes a capability-oriented view instead, formalizing four atomic operators — anchor, disentangle, and two others — that diverse multi-reference tasks share, so failures can be attributed to a specific operator rather than to a task name.

Leading multimodal models now reach roughly ninety percent on Video-MME, so conventional single-turn video understanding is saturating as a discriminator. VideoGAIA reframes the task agentically for general AI assistants, requiring tool use and multi-turn interaction with the video rather than one-shot question answering.

An Army memo indicates the service-wide Apache grounding ordered after a fatal crash in Texas will end this week, following completion of the directed inspections. The item is aviation-safety rather than AI, but it sits in the same fleet-management and readiness reporting pipeline that increasingly feeds predictive-maintenance models.

Video reasoning depends on evidence accumulated across many frames, so output-token-level on-policy distillation loses precisely the cross-frame structure that makes the teacher good. This work distills at the trajectory level in latent space instead, aiming to transfer the accumulation process rather than the final token distribution, in service of shrinking the compute cost of large multimodal video models.

Vision-language models can correctly identify every object and every attribute in a crowded scene and still attach an attribute to the wrong instance of the same class. Generic visual-question-answering accuracy just marks the answer wrong, while object-hallucination metrics see both object and attribute as image-supported, so neither reveals the transfer. InstaBind-Lite is a controlled benchmark that makes the failure directly measurable.

High-capacity encoders let the query dominate the latent state so that retrieved evidence becomes functionally irrelevant — a failure the authors name query dominance. GRIP imposes capacity asymmetry deliberately: the decoder keeps full-dimensional access to the query while retrieved evidence passes through a severe stochastic bottleneck, forcing the evidence channel to encode only what the query cannot supply.

This issue covers AI systems aimed at scientific work, a simulator built to study recursive self-improvement dynamics, and a reading of recent public comments framing large-scale technological pessimism. The recursive-self-improvement simulator is the item most likely to be reused: making the dynamics tractable in a sandbox is a precondition for arguing about them quantitatively rather than by intuition.

Retrieval-augmented systems apply a fixed trust policy toward retrieved evidence, which either over-trusts wrong context or under-uses context when the user explicitly asked for context-following behavior. Intent-Guided Decoding conditions the arbitration on inferred user intent, so 'answer from this document' and 'answer correctly' become distinguishable requests rather than the same decoding path.

Retrieval pipelines pre-materialize evidence through fixed chunking, embeddings, and persistent indexes, which commits to a granularity before the query is known and goes stale when files change. LENS formulates in-context search as budgeted evidence localization over a latent evidence space in the raw documents themselves, choosing granularity per query at search time. The trade is obvious — no index means more read cost per query — and the paper's contribution is making that trade explicit as a budget.

Building on prior work showing probes can extract user attributes such as age, gender, education, and socioeconomic status from model activations — and that intervening on those representations changes behavior — this post examines which textual cues drive the inference. The practical consequence is that a model may quietly filter options for a user based on an inferred attribute the user never disclosed and cannot see, which is an interpretability question with an immediate product surface.

A radiologist reading model output faces two failures: the model returns a number with no reason, and any system that turns that number into prose can quietly add claims the model never made. MIRROR chains a multi-label classifier, a Grad-CAM localizer that converts each positive finding into a named anatomical region, and a report writer that receives labels, probabilities, and regions but never sees the image. Withholding the image from the language layer is a structural rather than statistical guarantee against invented findings, which is the right kind of fix for this failure mode.

Unified image restoration built on pretrained latent text-to-image models inherits two liabilities: the variational autoencoder discards restoration-sensitive detail, and the open-ended synthesis prior invents content-inconsistent artifacts. PixRestore drops the autoencoder and operates a diffusion transformer directly in pixel space, which is the same latent-versus-pixel tension the large-scale training study above addresses from the pre-training side.

CUDA graphs remove kernel-launch overhead in principle, but capturing a real inference engine's workload means graphing as much as possible without losing compatibility with dynamic shapes, variable batch composition, and speculative decoding paths. The SGLang team walks through the techniques that extend graph coverage past the easy cases, which is where most of the remaining launch overhead in production serving actually lives.

A practitioner write-up reports a 33-percentage-point improvement in GPU cluster utilization with no hardware change, attributing it entirely to job ordering and scheduling policy. Fragmentation from mismatched job shapes is the usual culprit in shared training clusters, and the post is a concrete reminder that a meaningful fraction of the compute crunch is allocation rather than supply.

Large agents rely on retrieval to find the tools and APIs they need before they can plan or act, and this study shows that gate fails silently even with a fixed capability corpus: on ToolRet, a retriever fine-tuned on one source-specific slice collapses on another slice of the same benchmark, despite better aggregate scores. Fine-tuning a tool retriever on your own tool catalog therefore risks making it worse at everything else, which is a real deployment hazard for agents with heterogeneous tool sources.

Most code-generation systems use generated tests as static post-hoc validators, which cannot guide implementation and actively misleads when the tests themselves are wrong or incomplete. TDD-Agent operationalizes test-driven development instead, generating tests first and using them as the reasoning target during implementation, so test quality becomes an explicit intermediate artifact rather than a hidden failure mode.

Chronocooked is an Overcooked-inspired benchmark suite in which temporal information is unobserved yet critical for optimal performance, isolating implicit interval timing as a capability rather than confounding it with memory or planning. Baselines cover a non-recurrent model, a recurrent model, and a biologically plausible timing model, and the metrics are designed to expose where each fails. The motivation is human-robot interaction, where an agent that cannot represent 'how long since' or 'how long until' behaves in ways people read as inattentive.

Natural-language explanations are widely used as few-shot rationales, but which kind helps is unclear. Across six benchmarks and four instruction-tuned models, the authors compare explanation sources — human-written where available, self-generated, and generated by a stronger model — and relate downstream gain to faithfulness, separating whether an explanation helps because it is correct or because it imposes structure.

GitHub's write-up argues that a canvas — a persistent, inspectable surface showing an agent's plan, intermediate artifacts, and state — addresses three problems at once in hybrid human-agent development: opacity, lack of mid-run steering, and wasted tokens on work the human would have redirected. The cost argument is the least discussed and the most checkable: interruption is cheaper than rollback when a long agent run goes wrong early.

An updated survey of how working mathematicians are actually using AI systems, prompted by the recent run of headlines about models solving research-level problems. The consistent pattern is adoption concentrated in search-heavy interior steps — finding candidate constructions, checking cases, exploring parameterizations — rather than in problem selection or proof strategy, which lines up with today's matrix-multiplication result reaching a new bound by improving the optimization rather than the mathematics.

Tabular in-context learning models hit a context ceiling: beyond a few thousand rows, you cannot fit the training table in the prompt. Localizing the context with k-nearest-neighbor retrieval per query row lets the same model scale to larger tables by conditioning only on the locally relevant neighborhood, trading global structure for tractable context length.

Scientific figures and tables carry essential experimental evidence that digital libraries and multimodal systems still handle poorly. The ALD/E-ImageMiner benchmark and its ICDAR 2026 competition provide 1,951 figures from 205 publications with expert annotations for classification, data-table extraction, summarization, and visual question answering, targeting atomic layer deposition and etching specifically.

As video generation moves from single-shot clips to multi-shot narratives, the human character becomes the continuity anchor, but existing benchmarks measure appearance similarity or per-shot quality rather than whether physical and emotional state stay coherent across cuts. PersonaShot separates physical continuity, facial dynamics, and cinematic relations into criterion-specific evaluations, since each needs different visual, temporal, and relational evidence.

The family of training-free inference-time quality methods descended from classifier-free guidance was largely proposed on U-Net diffusion models and validated with metrics that score image quality in isolation, ignoring whether the image matches the prompt compositionally. Re-evaluating them on current architectures with alignment-aware metrics is exactly the kind of unglamorous replication work that tends to reorder which tricks people keep.

A discussion with the head of an independent evaluation company on how much weight regulation should place on benchmarks, given that a benchmark's value depends on what it measures, who builds and runs it, and how fast it evolves relative to the models. The relevant tension for policy is that benchmarks legible enough to write into rules are usually the ones most exposed to gaming, a point today's separate work on benchmark shortcuts in multi-agent systems illustrates directly.

New middleware lets LangChain agents pay for APIs under deterministic per-session budgets, signing x402 payments and tracing every transaction through LangSmith. Agent-initiated payment is the primitive that turns tool use into commerce, and the design choice worth noting is that the budget is enforced deterministically in the middleware rather than requested of the model — which is the only version of this that is safe to deploy.

Coding agents need business semantics, validation evidence, relations, and assurance that their context is current, which current systems externalize into retrieval indexes, summaries, graphs, rules, or reverse specifications that drift from the code. This proposal inverts that: selected code units directly carry agent-usable knowledge as source-bound executable knowledge units, so the knowledge cannot go stale relative to the code without breaking.

Serving LLM inference at radio-access-network edge sites means user sessions move between compute nodes as the user moves between cells, and a naive handoff discards the KV cache and re-prefills from scratch. Pallas migrates cache proactively ahead of the predicted handoff, trading a small amount of speculative network transfer for avoided recomputation.

Corrupted, inconsistent, or anomalous data quietly undermines medical AI, and regulators increasingly require dataset quality assurance for high-risk systems without specifying how to automate it. The authors build a controlled benchmark of seventeen realistic quality-assurance anomaly types across six categories in multi-center dynamic contrast-enhanced breast MRI, then evaluate unsupervised anomaly and out-of-distribution detection as the automated screen.