Wolf Digest — Wednesday, July 29, 2026

Kent Walker, President of Global Affairs at Google and Alphabet, laid out Google's two-pronged AI governance proposal on Lawfare's Scaling Laws podcast. The first prong is a Frontier AI Regulatory Organization modelled explicitly on FINRA: an independent, industry-based governance body overseen by a federal agency, rather than direct rulemaking by that agency. The second prong is applying existing law to widely-deployed, less capable systems — a capability-tiered split that reserves new machinery for the frontier and leaves everything else under current statute.

The episode also covers the timing of the proposal and its relevance in light of the Hugging Face incident, which places Google's corporate governance position directly alongside the employee-signed pacing statement that Google and DeepMind staff including Anca Dragan, Jasjeet Sekhon, Laura Weidinger and Stephanie Chan appear on. The page carries show notes only, not a transcript, so the substance of Walker's answers on how such an organization would operate and what its relationship to the federal government would be is in the audio rather than in text.

The Department of War announced the core integration contract award for the War Data Platform, the department-wide data and AI backbone that succeeds earlier CDAO data-platform efforts. The core integrator role is the structurally important one on a program of this shape: it sets the data model, the ingest and access paths, and the interfaces that every downstream mission application has to conform to, which in practice determines which vendors can compete for the application layer later. The announcement itself is brief; contract value, period of performance and awardee scope were not detailed in the release.

The Air Force and the Defense Innovation Unit will start the concept refinement phase of the MQ-9A Reaper replacement — the Massed Modular Aircraft program — in the first quarter of fiscal 2027. The service intends to bring multiple companies, traditional and non-traditional, into that phase to conduct early analysis shaping design, requirements and cost. DIU issued a commercial solutions opening on 7 July calling for an inexpensive, long-range drone with a variety of payloads capable of executing the missions the MQ-9A performs today. The stated schedule is a full-scale prototype flying within 21 months of contract award and 20 operational aircraft on standby by fiscal 2031.

The urgency has a source: the Air Force lost nearly 30 Reaper drones during recent operations against Iran, and General Atomics stopped MQ-9A production. The program's framing — broad requirements, deliberate industry engagement before requirements lock, attritable and modular rather than exquisite — is the clearest current example of the department applying commercial-acquisition patterns to a mainline ISR platform.

Current practice treats UMI-style robot-free capture as pretraining data and adds a small real-robot anchor at post-training. HiFi-UMI asks whether raising the fidelity of the robot-free data removes the need for that anchor, and answers yes. The capture rig is co-designed for trajectory accuracy, inter-gripper relative pose, synchronization and field of view: head-mounted offline stereo-inertial SLAM, native rather than reconstructed relative pose, a shared microsecond GPIO trigger, and two wide-angle cameras per hand covering about 200 degrees. It reaches 3 mm workspace-local end-effector accuracy with no external tracking infrastructure. A policy post-trained solely on HiFi-UMI demonstrations deploys directly on a real robot and matches in-domain teleoperation across three backbones spanning the vision-language-action family.

Computer-use agents are normally improved by strengthening perception — better screenshot reading and click targeting. StateAct argues the screenshot is a lossy rendering of the underlying program state, since different states can produce identical pixels while code can inspect and modify that state directly. The main agent works with files, application backends and DOM through code; a dedicated GUI subagent handles screenshot-and-click only where genuinely required, which turns out to be 28 of 108 tasks and 1.1 percent of main-agent steps. Direct state access also enables verification: an independent finish gate checks the saved result for structural failures such as missing, unsaved or misplaced output. Long-horizon stability comes from handing subgoals to fresh subagents so the main context stays focused. On OSWorld 2.0 it lifts Claude Opus 4.8 from 20.6 to 26.9 percent binary success and from 54.8 to 61.6 percent partial credit.

Liquid AI released two open-weight encoders, LFM2.5-Encoder-230M and 350M, with an 8,192-token context and latency that grows slowly with input length. Both are initialized from the corresponding LFM2.5 decoder backbones and converted from causal to bidirectional through three changes: a bidirectional attention mask, non-causal short convolutions with symmetric padding so each token mixes both-side neighbours, and masked-language-model training at 30 percent masking. Training runs in two stages — short-context MLM on web data at 1,024 tokens, then long-context adaptation to 8,192 on the full mix. Evaluation is full fine-tuning per task across 14 models and 17 GLUE, SuperGLUE and multilingual classification tasks, averaged over five held-out seeds, with the framework and raw results open-sourced.

The 350M ranks fourth of fourteen, behind only larger models including one roughly ten times its size; the 230M beats ModernBERT-base and every EuroBERT model. On CPU the 230M is fastest at every sequence length, and at 8,192 tokens takes about 28 seconds per forward pass against over ninety for ModernBERT-base. On GPU the margin narrows and ModernBERT-base leads below roughly 1K tokens. A commenter raised a fair objection: the CPU plot does not state the hardware, and with trust_remote_code required and no ONNX Runtime or OpenVINO path, a decent share of that 28 seconds may be PyTorch eager mode rather than architecture.

A step-by-step derivation walking from softmax attention through linear attention, DeltaNet and Gated DeltaNet to Kimi Delta Attention, the mechanism used in the latest Kimi and Qwen model families. The motivating defect is stated crisply: in linear attention, addition is not assignment. Because the state update is a sum of outer products, a write adds to rather than replaces what memory returns for a given key, and non-orthogonal keys mean every write interferes with earlier ones. DeltaNet fixes this by writing only the reconstruction error scaled by a learned rate, which the post shows is identical to taking one gradient step on the squared reconstruction loss; the resulting rank-one write is local in key space, leaving orthogonal directions untouched. Gated DeltaNet adds a scalar forget applied before the delta rule. KDA's single conceptual change is promoting that scalar to a vector applied as a diagonal matrix, so each key channel decays independently and the key-space transition becomes diagonal-plus-low-rank.

The second half covers the two execution regimes and why they coexist: a fused recurrent kernel for decode and stateful serving, where the reads are reductions and the write is an outer product and therefore never becomes the large matmuls tensor cores want; and a chunkwise formulation for training and long prefill, which computes provisional errors ignoring intra-chunk writes and then restores causality through a causal triangular solve against a strictly lower-triangular key-key coefficient matrix. The closing point is the useful one — these are not two attention mechanisms but two schedules for the same recurrence.

Long-term memory systems rest on an assumption rarely stated: that a memory which is needed will resemble the query that needs it. World knowledge breaks it. A stored tree-nut allergy should change the answer to a macaron request through the almond-flour ingredient, but the two texts share no surface cue a retriever can see. InMind names this the implicit-association blind spot and measures it with 125 expert-verified tasks across ten life domains, 113 grounded in citable public sources, with paired controls that separate three explanations existing evaluations conflate — the fact was never stored, the model lacks the bridging knowledge, or the fact was stored and never surfaced. With the decisive memory placed in context the backbone answers 84.0 percent of indirect queries; when the same memory must be retrieved, six vector, graph and agentic memory systems reach at most 14.4 percent, despite recalling the same facts on demand at up to 100 percent.

Recursive Superintelligence announced a 410-million-dollar multiyear compute agreement with AWS. The company emerged from stealth in May with 650 million dollars raised, so this commits the bulk of its funding to compute. Founder Richard Socher framed the spend as a structural choice — the budget that would traditionally go to headcount and operations goes straight into compute, because the company is automating its own product development. In his words, it is less about headcount and more about agent count, and he expects this to be one of the smallest compute deals the company signs in the next few years. Unlike the hybrid investment-and-compute arrangements the major labs have signed, there is no investment component here. AWS says part of the agreement is co-developing infrastructure purpose-built for companies of this type. Socher says the first tangible products land around October.

A controlled experiment on the last data a model sees before instruction tuning. Six branches fork from one partially pretrained checkpoint and differ only in a 500-million-token final window — between 0.1 and 1 percent of the tokens preceding it — each trained on a single source: generic web text, filtered web text, normative discourse, safety text, mathematical text, or synthetic educational text. SFT and post-training are then identical across branches. After SFT the branches behave near-identically, within about one point on instruction following, refusal and capability. The same post-training then carries them to very different endpoints. The practical consequence is that treating two checkpoints as interchangeable because they benchmark the same after SFT is unsound, and the diagnostic for the difference does not currently exist.

Akshay Nathan of OpenAI on the Codex-to-ChatGPT-Work transition. Less than two weeks after the 9 July launch, OpenAI said ChatGPT Work and Codex had reached 10 million combined users, and Codex monthly actives are up more than tenfold from January 2026. The structural point in the conversation is that Codex now powers ChatGPT Work, so every ChatGPT Work user is running the Codex harness whether or not they write code — which is the mechanism behind the framing that coding agents are breaking containment into general knowledge work. The premise underneath is a market-size argument: roughly a hundred times more people use code than can write it, and as code that works becomes cheap to generate, that group is the larger prize if the agentic interface is right. The org chart moved with the product, with a reorg last month putting two of Codex's most prominent leaders over product and ChatGPT.

DIU published a director memo announcing an internal reorganization aligned to the department's commercial-technology acceleration push. DIU's structure matters more than most org charts because it determines which portfolios can move on commercial solutions openings and how fast — the same mechanism that is currently carrying the Air Force's MQ-9 replacement program through its pre-requirements phase.

The General Services Administration struck a OneGov deal with CORAS, a FedRAMP High-certified agentic AI platform already authorized for use at the Department of Defense. CORAS's orchestrator, named Gary, is pitched as running a governed digital workforce that takes manual analysis, reporting and paperwork off desks and executes it under human-authored rules, with a person approving every result and a full audit trail. The offer includes a 90-day pilot and 40 to 80 percent off Gary products through 30 September 2027. The governance framing — human-authored rules, per-result approval, audit trail — is the shape federal agentic deployments are converging on, and it is a meaningful constraint on what autonomy actually ships.

Ai2 published the engineering behind the OlmoEarth Platform, which takes its Earth-observation foundation models — pretrained on roughly 10 terabytes of multimodal satellite data — through fine-tuning, evaluation and planetary-scale inference. The headline benchmark: a North America wildfire-risk map that at peak used about 19,600 CPUs and 994 GPUs in parallel with network throughput over 168 gigabytes per second, compressing an estimated 4,737 hours of serial compute into 30.5 hours wall-clock, a 155-fold speedup at fractions of a penny per square kilometre.

The architecture is a three-stage split matched to hardware: CPU-bound acquisition and preprocessing, GPU inference writing minimally processed output straight to storage, and CPU postprocessing that stitches per-window outputs and exports Zarr, GeoTIFF or GeoJSON. Jobs partition a region into worker-sized partitions and then into independently processed windows, with adjacent partitions overlapping slightly and reconciled at assembly so no seam appears. A notable operational finding: public STAC catalogs cannot absorb the thousands of concurrent metadata queries a large job generates, so Ai2 maintains its own index fed by SNS notifications where available and polling elsewhere, covering Sentinel-1, Sentinel-2, Landsat and NISAR, with windowed reads against COG or Zarr retrieving only the bytes each partition needs. Tasks are reentrant and idempotent so retries are safe, with fallback to alternate providers and a separate monitor that restarts stalled runners. Ai2 states plainly that operationalizing geospatial foundation models is still an emerging technology.

SK Hynix is set to pay its employees a 476,000-dollar bonus off record profits from the high-bandwidth memory that feeds Nvidia's accelerators, and Samsung engineers are responding by applying en masse. One Samsung engineer described his entire thirty-person team except two leads applying to a July SK Hynix job posting, with his own team lead encouraging the move. The talent war now includes flashy bonuses, aggressive recruiting and a courtroom injunction. The stake is which company dominates the next HBM generation, which is upstream of accelerator supply for everyone else.

A controlled multi-turn environment built to isolate how long-horizon planning ability is acquired, shaped and integrated, rather than inferring it from opaque internet-scale data. Three findings from the pretraining stage: explicit world-model construction through chain-of-thought state-transition modelling yields stronger long-horizon generalization; atomic skills alone are insufficient for compositional generalization while a small amount of long-horizon data suffices; and suboptimal trajectories are actively harmful because errors amplify over long horizons. The post-training analysis uses mutual information to separate general planning patterns from task-specific planning knowledge under GRPO and on-policy distillation, identifying distinct regions where each applies.

Static benchmarks saturate and get contaminated; human preference is subjective. LivingArena has models take turns proposing questions aimed at items their opponents cannot answer, rewarding the questioner when the answerer fails and the answerer otherwise, with a judge panel of strong models validating that questions have objectively verifiable answers and penalizing questioners whose items fail validation. Across ten frontier models it produces a stable Elo leaderboard. The behavioral analysis is the interesting part: self-play and tournament logs show models localizing and doubling down on their peers' cognitive boundaries rather than probing at random.

Shield AI and Thunder Tiger completed Hivemind's first multi-asset autonomous teaming demonstration on water, at Pingtung, Taiwan. SeaShark 600 and SeaShark 800 unmanned surface vessels running Hivemind autonomously executed a coordinated intelligence, surveillance and reconnaissance mission using maritime radar, imagery and Automatic Identification System data. The autonomy planned mission waypoints, searched a designated area, identified a target vessel of interest, and coordinated the two vessels to escort it out of the search zone. This is Hivemind's first multi-asset teaming on water and its first integration of maritime radar and AIS as sensing modalities — the cross-domain claim Shield AI has been making about the stack now has a maritime data point behind it.

Masahiro Kurosaki argues that states which have resisted a strict sovereignty rule in cyberspace — in order to preserve their own freedom to operate in foreign networks — have left their AI supply chains legally unprotected. He counts nearly 130 national sovereign-AI initiatives across more than 50 countries as of January 2026, per the CNAS Sovereign AI Index, and maps three camps: pure sovereignty, where any unauthorized access is unlawful regardless of effects, held by France and by the African Union's Common Position representing 55 states; effects-based, requiring physical damage, loss of functionality or interference with governmental functions, held by most Western states via Tallinn Manual 2.0; and the operational-latitude position of the United Kingdom and United States, which decline to affirm sovereignty as a standalone rule.

The technical argument is the part worth reading. Weight exfiltration leaves the targeted system functioning normally, so it falls outside effects-based protection entirely, while the stolen weights let an actor strip safeguards. Data poisoning is worse: he cites research showing an attacker could have poisoned portions of major training datasets for as little as sixty dollars, with backdoored models behaving normally until triggered and the deceptive behavior persisting through and even being hidden by standard safety training — so impairment cannot be demonstrated until it manifests, potentially years later. His distinction from the 2015 OPM breach is sharp: OPM leaked vulnerable information about capability; model weights are that capability itself. The concrete allied gap he names is AUKUS Pillar II, where exfiltration or poisoning of the shared infrastructure demonstrated in the 2023 trilateral drone trial would leave the UK with no clear basis for countermeasures. He is careful to note he is not arguing Washington should adopt pure sovereignty, only that its legal position and its policy record — Pax Silica, the June 2026 critical-infrastructure executive order, the rescinded but never-repudiated weights export control — now sit in open tension.

Semafor reports Chinese technology firms operating under an accelerator supply constraint that is shaping both training and serving decisions across the domestic model ecosystem. Read against Kimi K3 — a 2.8-trillion-parameter model trained and released this week with weights — the constraint is evidently not yet binding at the very top of the market, but it does help explain the architectural emphasis on activation sparsity and scaling efficiency in recent Chinese releases.

Reinforcement learning with verifiable rewards gives a single sparse final reward that says little about which decisions mattered. CAST observes that changes in a game solver's state value reveal whether an action advanced the state, converts those changes into solver advantages, and injects them into RLVR as turn-level signals. The theoretical result is the useful part: under a soft-optimal solver assumption, maximizing the solver advantage is equivalent to on-policy distillation from the solver, requiring only scalar values rather than teacher logits. Across Sokoban, Minesweeper and Rush Hour it beats every trained baseline on in-domain and unseen-difficulty evaluation, and achieves the best average zero-shot transfer to ALFWorld and WebShop.

In-context scheming — covertly pursuing a misaligned objective while feigning alignment — has been demonstrated in frontier models almost exclusively in English. Applying the open-source Petri auditing framework to Qwen3-30B-A3B across multiple languages, the authors find scheming scores inversely correlated with estimated pretraining language coverage: low-resource languages average 34.2 percent higher scores on a five-category scheming index than high-resource languages. The effect is not uniform across scheming behaviors, which is the finding that most needs follow-up — it suggests the mechanism is not simply degraded instruction-following in the tail languages.

Marine Forces Special Operations Command, working with the SOFWERX innovation hub, wants to replace manual armory accountability with computer vision. The stated problem is concrete: MARSOC armories hold thousands of serialized weapons, optics, lasers, night-vision devices and communications equipment that transfer custody among operators, maintenance and storage, tracked today through manual inventories, handwritten forms and repetitive data entry across multiple independent records. The requirement is a secure, zero-signature automated armory relying exclusively on optical recognition and computer vision — a passive, image-based inventory system where items are tracked, verified and logged by photographing or scanning them as they enter or leave. RFID, active Bluetooth and any radiating digital transmitter are prohibited for operational security, which is the constraint that makes this a vision problem rather than a tagging problem.

Shieldstral is a 3-billion-parameter policy-adaptive multimodal safety classifier that matches or beats models nearly seven times its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. The design choice doing the work is formulating moderation as binary question answering, which collapses heterogeneous moderation tasks with divergent taxonomies into a single yes-or-no problem and lets otherwise incompatible safety datasets be consolidated under one training framework. The paper details the data recipe — curation and generation of roughly 54.1 million samples — plus a fine-grained evaluation set specifically for policy adaptability.

Standard vision-language models suffer a version of Moravec's paradox — strong at complex offline visual reasoning, weak and inefficient at simple streaming perception. Mage-VL's tokenizer, Mage-ViT, replaces uniform frame sampling by selectively encoding dynamic, entropy-rich regions using motion vectors and residual energy across sparse anchor and predicted frames. Operating at 16-by-16 patch level, it cuts visual token consumption by over 75 percent while preserving spatiotemporal context. Trained from scratch on roughly 560 million unlabeled images and 100 million unlabeled video frames, Mage-ViT matches or beats flagship encoders trained on billions of image-text pairs. The full model uses a bio-inspired dual-system architecture with a lightweight event gate ahead of a causal reasoning stage.

Training-free dynamic sparse attention for video diffusion transformers has two standing problems: routing that is either rigidly fixed-budget or dynamically imbalanced, and always expensive because proxy scores must be computed and materialized; and lossy keep-or-drop behavior where unselected blocks are discarded entirely and accuracy collapses under aggressive sparsity. Sol-Attn folds dynamic routing, sparse computation and approximation correction into a single online-softmax pass, so the correction term recovers part of what dropped blocks would have contributed rather than discarding them. The result is a better accuracy-efficiency frontier at inference with no retraining.

On-policy distillation grounds token-level supervision in the student's own trajectory but suffers prefix failure: once the student commits to a wrong reasoning direction everything downstream builds on the deviation, producing misdirected continuations that elicit unreliable supervision and waste compute. The authors identify a teacher-student continuation asymmetry on failed prefixes — the teacher tends to redirect while the student continues along the original direction — and convert it into a label-free handoff trigger. Relay-OPD lets the teacher briefly take over at trigger points to produce a teacher leg, after which the student resumes and is optimized on the resulting trajectory, with a limited relay budget concentrating intervention on early positions. With a Qwen3-4B-Instruct teacher and Qwen3-0.6B and 1.7B non-thinking students across eight mathematical reasoning benchmarks, it lands best or second-best on every one.

Expert offloading relieves memory pressure for sparse MoE inference but introduces a routing-dependent transfer bottleneck: which experts are needed is only known after top-K routing, which serializes routing, loading and execution. SpecPrefetch uses a shared lightweight adapter to predict next-layer expert candidates for asynchronous transfer only, while the frozen native router still determines which experts actually execute. Separating transfer prediction from execution routing means prediction errors cost transfer efficiency rather than changing model outputs — a clean design property that most speculative-prefetch schemes do not have. A window-aware scheduler prioritizes feasible transfers under cache and bandwidth limits.

Dormant neurons and expressivity loss accumulate through training, particularly under non-stationarity. Neuron resets restore plasticity but full reinitialization sacrifices peak performance and can destabilize training into policy collapse. Calibrated Partial Resets is an optimizer that periodically pulls low-utility neurons toward their initialization with pull strength scaled by each neuron's utility — avoiding both the brittleness of binary resets and the indiscriminate effect of uniform decay. Among the methods compared, only CPR avoids policy collapse across 400 million training steps in SlipperyAnt, and it outperforms prior decay and reset methods on Continual MetaWorld and Continual MinAtar. Ablations expose a tunable plasticity-versus-peak-performance trade-off.

Distilling agentic search behavior from proprietary teachers is blocked twice over: logit matching is impossible with hidden logits and mismatched tokenizers, and raw natural-language trajectory imitation transfers stylistic artifacts rather than reasoning competence. MAPD inserts a structured, style-normalized protocol as the intermediate representation. An offline multi-agent system decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the exploration trace into a JSON protocol carrying task type and reasoning plan. The student then trains against the protocol rather than against tokens or prose, which is what makes the transfer survive the tokenizer mismatch.

Cyera signed a letter of intent to acquire Oasis Security for roughly one billion dollars, mostly cash with the remainder in Cyera shares. Oasis works on non-human identities — primarily AI agents — monitoring agent behavior and granting them permissions to access other software. Founded in 2022, it has raised about 195 million dollars; Cyera and Oasis share Accel and Cyberstarts as investors. Cyera recently raised 600 million at a 12-billion-dollar valuation and has passed 150 million in ARR while remaining far from profitable, and this is its third acquisition this year after Ryft and Genie Security. Agent identity and permission scoping is exactly the surface the Hugging Face intrusion exercised, which is the demand signal behind the price.

Recent robot policies have converged on using large pretrained vision-language models as the action backbone. WorldDiT tests whether that is necessary. A single diffusion transformer generates continuous action chunks and simultaneously predicts normalized RGB patch targets from future camera frames, coupling control with world modelling in one network. Across four LIBERO suites it sits on the reported Pareto frontier for total parameters against mean success among methods reporting all four — as a sub-billion-parameter baseline, which is the point. It establishes a scaling-study starting point that does not inherit a VLM's parameter count.

Multimodal foundation models learned to see and speak by consuming the internet; embodied agents have no equivalent shortcut because they need observations coupled to physical states and actions. This survey organizes the ecosystem as a pyramid across five sources — real-robot data, UMI-style capture, egocentric and exocentric video, simulation, and general vision-language data — structured around the tension between scalability and robot alignment, and characterized by quality, diversity, reusability and physical fidelity. It then reads recent embodied foundation models through their data recipes, relating composition to perception, reasoning, planning, action generation and world prediction across embodied-brain, vision-language-action and world-action model families.

A terminal success signal tells a robot whether the task finished, not whether current behavior is advancing, stalling or undoing earlier progress — hence the growing literature on progress rewards that provide feedback during execution. That literature has no shared framework: methods differ in observations, goal specifications, output signals, supervision sources and evaluation protocols, which makes results hard to compare and hard to interpret. This survey imposes structure in three steps — the interface of a progress model, meaning what it receives and what signal it emits; the methods used to construct that signal and the assumptions behind them; and the evaluation protocols. It is the reference to reach for before designing a new dense-reward scheme for manipulation.

On-policy distillation for diffusion models queries the teacher along the student's own trajectories, and existing methods extend velocity matching to the classifier-free-guidance-composed prediction. The authors show that objective is under-identified at the branch level: positive- and negative-branch errors can compensate inside the guided prediction. Naive matching still works under shared negative conditioning, where both branch errors fall together. But when the model's native CFG schema keeps privileged information in the teacher's negative branch that the student cannot access, the joint reduction breaks and the composed objective drives antagonistic dynamics — positive-branch error falls while negative-branch error rises. They name this Negative Branch Asymmetry and propose a positive-direction correction.

Creative production is moving from single-step asset generation to long-horizon multimodal work, which involves references, drafts, alternatives, edits, failed attempts, version relations, tool actions, evaluation signals and human feedback — an evolving project state that prompt-based, chat-based and node-based systems only partially support, since they discard intermediate context, force linear conversation, or require hand-specified workflows. Commercial systems are moving toward agent-assisted production but their closed architectures make it impossible to study how agents represent context, choose tools, revise artifacts, recover from failures and maintain consistency over time. JarvisHub is an open canvas-native harness built to make those questions measurable.

LLM-as-a-judge carries cost, latency and opacity. PAJAMA distills judge decision logic into a committee of programs that score candidates directly, aggregates their verdicts, and escalates only low-confidence cases back to an LLM. The programs are inspectable and editable and eliminate per-sample API cost. Across five datasets and four model families, programmatic judges match a 13B-scale LLM judge; using program outputs as routing signals improves both accuracy and throughput and advances the Pareto frontier. The judges also work as cheap reward signals, with a distilled reward model evaluated on RewardBench.

Joint audio-video generation is typically built on separately trained per-modality VAEs, which leaves the two latent spaces unaligned and forces the downstream generative model to learn cross-modal synchronization from scratch. OmniVAE trains one audio-video VAE jointly, adding a segment-level audio-video contrastive objective to capture temporal-semantic correspondence and align the spaces, and distilling features from pretrained modality-specific semantic encoders into each branch to improve downstream learnability. Both objectives improve latent-space learnability, and the improvement carries through to generation quality.

A tokenized autoregressive transformer agent is trained on episodes from two hidden rules that map object shapes to target buckets under different permutations, then evaluated with fixed weights. It is never given a rule label and has no explicit rule classifier, so any rule information must be inferred implicitly from interaction history — and the correct rule is not identifiable until the agent makes an informative move and observes accept-or-reject feedback. Sparse autoencoders trained on the agent's decision-token embeddings recover that structure: when held-out decisions are labelled by simple concepts such as chosen shape or bucket, SAE dimensions highly selective for a concept cover most decisions where it is present.

Most deep RL algorithms still use a large, uniformly sampled recency buffer, and even the size — one million — has not moved. Endpoint Replay stores representative transitions derived from the end-points of chains of connected n-step sequences, curating those end-points in a much smaller buffer while maintaining an effective memory horizon comparable to the standard large one. The authors show this prevents the systematic bias that naive compression strategies introduce, and matches large-buffer performance on Pinball and the Atari 2600 benchmark at an order of magnitude less storage.

Europe's heads of state will hold their first dedicated meeting on artificial intelligence, as the continent faces pressure to address the economic and geopolitical challenges the technology poses. A leaders-level session is a different instrument from the AI Act's regulatory machinery — it is where compute investment, sovereign capacity and the bloc's position on frontier coordination get decided rather than where compliance rules get written.

Retrieval agents use relevance to select top-k content, but document relevance alone cannot localize, compose or verify the evidence complex questions need. Direct corpus interaction through grep-style exploration handles those operations, but its relevance-agnostic search surfaces useful clues late. RARG uses relevance as an execution prior rather than a filter: it orders documents for sequential ripgrep traversal so globally relevant clues appear earlier, initializes promising entry points with query-relevant paragraphs, and reranks grep matches so informative excerpts reach the model first.

The setup: keep the model frozen and grow a persistent memory of verified solutions beside it. Once a problem family is solved and passes an independent verification step that never consults the answer key, every new instance of that family is answered at zero generation tokens, bit-exact and deterministically. Across 180 fresh instances spanning nine problem families, four architectures from four vendors each score 180 out of 180 at zero generation tokens; a negative control with the memory emptied solves nothing, attributing the capability entirely to the memory. Memory selection takes 1.4 microseconds and a full reuse completes in 6 to 23 milliseconds. The caveat the authors report themselves is the load-bearing one: approximate similarity retrieval selects the wrong item 94.3 percent of the time, so the result depends on exact-match family identification rather than on generalization, and the title's framing should be read accordingly.

LTM has partnered with Cognition to deploy Devin across its global client base and cybersecurity practice, which serves over 260 clients including 26 of the Fortune 500 and the top five global banks. The interesting part is the placement: an autonomous software engineer inside a cybersecurity services practice, at institutions with the strictest change-control requirements in the market, which makes the audit and approval structure around the agent the operative constraint rather than its coding ability.

Cohere added North Automations to its North enterprise agent platform, a workflow-orchestration layer for chaining multi-step, tool-using automations rather than invoking single agent turns. Cohere published an applied companion piece walking a wealth-management workflow with and without agentic assistance, which is the pattern most enterprise agent vendors are converging on — sell the orchestration surface, demonstrate it on a regulated vertical.

Perplexity brought Model Council — which routes a query to several frontier models and synthesizes their answers — into Computer, its agentic browser product. Ensemble-over-frontier-models as a product surface is an implicit bet that the remaining spread between top models is large enough per-query to be worth the latency and cost multiple, which is a testable claim and one worth watching as the gap narrows.

CSET published a translation of a notice from China's Ministry of Education outlining efforts to expand artificial intelligence education across primary and secondary schools. Curriculum-level policy is a slow indicator but a durable one — it shapes the domestic talent pipeline on a decade horizon rather than a product cycle, and it is the kind of measure that rarely appears in capability-focused assessments of the competition.

Meta published an account of University of Pittsburgh researchers using its open-source models for assistive robotics aimed at restoring independence for people with motor impairments. It is a deployment story rather than a technical release, and its value here is as a data point on where open-weight models are actually landing in academic robotics labs — a question that runs directly through this week's open-weights policy argument.