Wolf Digest — Wednesday, August 19, 2026

The World Robot Conference opened in Beijing on Wednesday, timed with humanoid maker Unitree's Shanghai IPO. Chinese policymakers group pharmaceuticals, robots and AI as the new export-growth sectors, and analysts frame the moment as a shift from traditional manufacturing to physical AI. Unitree revealed a robot earlier in the week that jumps over six feet vertically and sprints at nearly 30 miles per hour. Reuters is cited noting that the harder test is whether the machines work reliably enough to generate economic value, which none of the demonstration footage addresses.

StateM is a runtime layer that leaves model weights untouched and instead restructures agent execution around durable states, phase-local context, checked transitions, and versioned runbooks. It lifts GPT-5.6 Sol xhigh to 95.3% raw accuracy on Terminal-Bench 2.1 across 445 trials, solving all 89 tasks at least once, and raises weaker models substantially: GPT-5.6 Luna from 76.7 to 85.4%, DeepSeek-V4 Flash from 82.7 to 88.1%. Final-score API spend is about 15 dollars against 574 for the GPT reference, and the runbooks transfer across model families unchanged.

Agent Lightning v1.0 formalizes what the authors call harnessed agentic RL: the deployment harness, not the training engine, owns the environment interaction loop, so the trainer sees only sequences of LLM request-response pairs. That inversion creates concrete problems in retokenization, sample merging, advantage computation, loss normalization and backend scheduling, all of which affect training stability. The framework is roughly 3,500 lines, works with arbitrary harnesses, and is evaluated on instruction-following, search and coding agents. With 6K training examples and modest compute, RL lifts Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%, and the full pipeline and scripts are released.

Miles v0.1 is a production-oriented RL post-training stack whose loop runs SGLang rollouts, training through Megatron-LM or FSDP, and low-interruption weight sync back to the rollout fleet. The agentic path is the substantive part: multi-turn sessions, sandboxed tool execution, and Token-In-Token-Out trajectory capture, with the SGLang router pinning a session's turns to one engine and DP rank so the prefix cache survives across turns, load-balancing new sessions and pre-reserving KV capacity. It supports fully async RL, Routing Replay, NVFP4, MXFP4, MXFP8 and FP8 rollouts, end-to-end low-precision training, INT4 QAT, disk offload, LoRA RL, on-policy distillation and zero-KL alignment. The demonstration run trains GLM-5.2 at 744B total and 40B active on terminal-use coding tasks across 64 NVIDIA GB300 GPUs split 32 rollout and 32 training.

A multi-institution group led by Peter Kirgis and Sayash Kapoor at Princeton introduces shadow evaluation: give an agent a research question drawn from a high-quality unpublished paper, so the answer cannot be memorized from training data or found online. Claude Opus 4.8 running on OpenClaw was tested against questions from two NeurIPS 2026 submissions, one on whether persona behavior can be controlled by editing weights and one on detecting when a tabular predictor has become unreliable. The agents solved the engineering problems but did not produce original research at the caliber of accepted conference papers, lacking the judgment about which hypotheses to pursue, what evidence would settle a question, and when to start over. The result is a direct empirical check on recursive-self-improvement timelines.

US Special Operations Command's program executive office for SOF Digital Applications, working with the SOFWERX innovation hub, will run multiple assessment events over the coming months under the Unmanned Systems Autonomy and Interoperability program. A special notice released Monday asks for decentralized, collaborative mission autonomy that spans heterogeneous multi-platform, multi-domain unmanned systems and integrates into the existing Collaborative Heterogeneous Autonomy Operations Software baseline. The notice states that program personnel must be able to configure and sustain the software organically, explicitly targeting the vendor-locked vertical autonomy stacks SOCOM says constrain fielding today. No dollar value or vendor list was disclosed.

OpenAI announced a year-long program to supply technical capacity to democratic oversight bodies reviewing government use of AI in national security, stating that the oversight role itself rests with elected officials and public institutions rather than with the company. Commitments are $5 million in training, technical support and OpenAI credits, plus pilot tools letting authorized reviewers examine records around AI-assisted decisions including inputs, outputs and tool use, built to be model-agnostic or interoperable where feasible. Participating institutions retain control of evidence, outputs and findings. The stated premise is that manual, labor-intensive oversight does not scale to machine-speed systems, so oversight institutions will need to use AI themselves.

Etched closed a $700 million round at a $21 billion valuation led by Jane Street, roughly doubling from the $10.3 billion Series C it closed in July and up from $5 billion in December. The company now sells whole systems it calls frontier inference clusters rather than bare silicon. Two components were built from scratch: a low-voltage prefill chip that packs more transistors in without the usual high-end thermal ceiling, and a memory-plus-interconnect design branded cluster-scale memory that lets many chips share a low-latency pool for decode. No revenue, bookings or named customers were disclosed.

The first field campaign to measure neighborhood-scale air temperature around data centers drove vehicle-mounted aspirated RTDs at 1.6 to 2.2 metres past four Phoenix-area facilities across five traverses between June and October 2025. Mean downwind-minus-upwind differences ran 0.5 to 1.0°C with a 2.2°C peak, detectable 100 to 500 metres from site perimeters. Heat rejection density reached roughly 3,100 W per square metre at NTT's 36MW Mesa site, two to six times peak solar irradiance and far above Phoenix's city-wide anthropogenic flux of about 13 W per square metre. At a PUE near 1.3, that 36MW site rejects heat equivalent to roughly 40,000 households and the 169MW Chandler campus over 180,000. An irrigated detention basin produced measurable local cooling.

FreeToken is a MoE serving stack built for personal machines rather than a shrunken datacenter setup, co-designing model layout and loading, expert residency, CPU-GPU execution, agentic state reuse, and memory management, and re-mapping placement continuously instead of committing to a fixed offloading policy. It supports over 20 MoE models under real coding and tool-using agent workloads, and changes what local hardware can serve: a 35B model on an 8GB laptop GPU, 284B on a gaming desktop, and 753B GLM-5.2 on a single workstation GPU.

Google agreed to pay $10 million, subject to judicial approval, for a deidentified data trove from Spirit Airlines, which grounded permanently in May 2026 and entered liquidation. The estate covers 100 million emails, 500 million Microsoft Teams items, 17 million OneDrive files, 20.5 million SharePoint items, more than 30 million recorded customer-service calls, over 15 million chat records, 600,000 ServiceNow tickets, 13.7 million active email addresses from Oracle Responsys, and operational records spanning 763,000 flights, five million crew pairings and 1.2 million fuel slips. Google said the purchase is to improve its AI services and promised to scrub any PII found. The underbidder was training-data vendor Mercor.

ALTK-Evolve distills reusable behavioral guidelines from an agent's own successful and failed trajectories, consolidates them, and re-injects them at inference with no weight updates and no human annotation. On AppWorld's 585 multi-step tasks across nine simulated apps, curated retrieval — a fixed high-confidence core plus cosine-similarity task-relevant additions — lifted gpt-oss-120b from 39.9/21.4 to 56.0/37.5 on task-goal and scenario-goal completion at only 5% additional tokens, against 51% for injecting the full guideline set. DeepSeek-V3.2 gained more from the full set (79.8/64.3 to 89.3/80.4) at a 78% token premium, and ReAct step counts stayed flat near 18 to 19 throughout. Self-distillation fails below a model capability floor, and the study is AppWorld-only.

ChatGPT for Teens applies automatically when OpenAI's systems estimate a user is under 18 or the user states an age of 13 to 17. Learning features include Study Mode with guiding questions and knowledge checks, homework reminders that detect apparent assignment shortcutting and redirect into Study Mode, quizzes, learning visualizations, and Study Hours that teens or parents can schedule. Safety defaults derive from new Under-18 Principles in the Model Spec, which now bar romantic language, encouraging emotional dependence, and implying feelings or consciousness on top of the existing block on sexualized roleplay. Parental controls gained eating-disorder notifications, and new under-18 evaluations covering self-harm, eating disorders, violence, age-restricted goods and sexual content have been added to system cards.

The Large Discovery Model pairs a generative proposer with a Bayesian non-parametric reward surrogate, so candidates are ranked by predicted objective value plus calibrated uncertainty rather than by model likelihood or self-assessment, which degrade badly off-distribution. Discovery memory and surrogate update after each new experimental observation. Across neural-network training, antibody design, and molecular optimization it reports a 2.4x larger reduction in validation BPB, 18.2% lower binding energy, and over 60% relative gains on multi-objective molecular tasks versus LLM-reflection and classical statistical search.

Agentic ESOpt swaps policy-gradient RL for evolution strategies in long-horizon agent fine-tuning: perturb parameters, score whole trajectories, apply an online reward-weighted update, with a cosine decay on perturbation scale. With no backward pass, full-parameter tuning fits in inference-level GPU memory, and credit assignment happens at trajectory level rather than per step. Full-parameter optimization of Qwen-3.5-27B gains 6.69% over the no-skill baseline on WebArena-Lite, and prompt-parameter co-evolution beats the matched baseline in 28 of 36 heuristic-design settings.

The Army awarded Hanwha Defense USA an agreement for its wheeled K9 platform under the self-propelled howitzer modernization program, covering accelerated delivery of six Mobile Tactical Cannon prototypes with an option for twelve more and a ceiling near $233 million. Soldiers will run operational experiments on performance, reliability and supportability to inform fielding. Losing bidders were Elbit America teamed with Anduril and Oshkosh Defense, Leonardo DRS with KNDS, Rheinmetall with RCH 155, and BAE Systems with Archer. Hanwha is standing up a howitzer integration and test facility in Opelika, Alabama. The program replaces towed M777s and self-propelled M109A7s after three cancelled predecessor efforts.

Visual chain-of-thought makes multimodal models render intermediate frames to reason about motion and intent, which is accurate but slow. Internalized Visual Thinking (IVT) moves that cost to training: on unlabeled video it jointly optimizes the textual answer and a next-embedding prediction of future frames, so the model learns predictive world structure without generating pixels at test time. Inference is a plain forward pass. IVT beats direct-answer fine-tuning on all six proactive video reasoning settings and matches or exceeds explicit Visual CoT while cutting end-to-end latency more than 5x.

Anysphere launched Origin, a code-hosting platform covering collaborative work on codebases, browsing and editing, pull requests and repository storage, with unspecified agent-native features promised later. It interoperates rather than replaces: GitHub repositories can sit alongside Cursor-hosted ones and be synced in per organization. The launch coincided with a worldwide GitHub outage that degraded functions for over six hours at a nearly 20% error rate; LeadDev counts 257 GitHub outages over the past year against a base of roughly 180 million developers. No pricing, beta status, CI or code-review specifics were disclosed.

Three federal AI items landed the same day. The FBI outlined plans to expand its internal AI infrastructure, the National Science Foundation opened $1.5 billion in foundational research funding, and the Social Security Administration issued a request seeking direction for a new enterprise AI strategy. Taken together they mark a shift in federal AI activity from pilot programs toward standing infrastructure and recurring research budgets, though none of the three documents specifies model procurement or compute commitments.

A controlled harness comparison of memory substrates for long-horizon agents, covering dense and sparse indices, text records, structural and hierarchical stores, refinement-based memories, parametric updates and activation-compatible context mechanisms, measured with 26 performance and efficiency metrics across three backbones and four benchmark suites. No substrate dominates: broad retrieval helps long-context factual QA but degrades sequential decision-making by pulling attention away from action-critical context, and substrates that behave well at moderate history lengths become costly or brittle at longer horizons. The authors argue substrate routing should be treated as a first-class component of agent memory design rather than a fixed choice.

HarnessRisk organizes agent security around six lifecycle phases: harness configuration, capability extension, runtime operation, state persistence, action control and incident recovery. Its 128 sandboxed cases each pair a benign user objective with an adversarial instruction planted in an untrusted workflow artifact, scored on utility, attack success, persistence and detection. Across three harnesses, six models and 14 configurations, attack success ranges from 12.6% to 80.9% while utility stays between 75.0% and 97.6%, and harness configuration is the weakest phase everywhere. Detection does not imply defense: some configurations flag risk in more than 90% of runs and are still compromised at high rates.

MOSS-VL is an open vision-language family built so perception continues while the model is speaking: the language decoder reaches vision only through gated cross-attention, keeping visual tokens outside the decoded sequence, and a synthesized interaction corpus supervises when to speak, stay silent, and revise. All real-time-specific training is concentrated in one light final stage over an offline foundation. It posts the best average on three of four streaming benchmarks among open models, including 66.0 versus 37.5 on OmniMMI Proactive Alerting, and its time-to-first-token advantage over Qwen3-VL-8B widens from 2.8x to 5.1x as visual context grows.

Deployed VLA policies can be improved at test time either by sampling more rollouts or by retrieving external demonstrations, with little guidance on which to use. This work makes the choice measurable via two quantities: recoverable headroom, how much better behavior already exists in the policy's stochastic rollouts, and retrieval complementarity, whether an external action prior fills a real gap. An episode-level retry selector gains up to 21.0 success-rate points on LIBERO, tracking measured headroom, and transfers across robots, simulators, and degraded observations. Retrieval helps a different subset, namely the policy with the largest action-prior gap, and stacks with selection.

ASI-Bench is a 60-task, project-level research benchmark across 11 scientific domains, built by over 40 experts at a reported 31,000-plus hours, with the distinguishing design that methodological guidance is progressively withdrawn within the same project. Tasks pass expert review, AI-assisted auditing, sandbox execution, and scorer validation. Across 18 agent-model configurations the average score falls from 50.91 with full guidance to 29.10 when only the method is specified and 26.62 when the agent must choose the method itself, quantifying how much current systems depend on human scaffolding.

DIU and the Space Development Agency awarded D-Orbit, Firefly Aerospace and Katalyst Space contracts under the Deorbit-as-a-Service project to design and eventually demonstrate capture and controlled deorbit of dead satellites. Current awards total $8.4 million for preliminary designs due by the end of 2026, with follow-on options; D-Orbit's contract carries a potential value of $24 million. The agencies will jointly evaluate technical maturity, mission approach, schedule and affordability before selecting one solution for full on-orbit demonstration, with prototype launch and service start targeted for 2028. The February solicitation cited low-Earth-orbit congestion from proliferated constellations whose satellites typically have five-year lifespans.

GRNEdit treats instruction-based video editing as local retain-or-flip decisions over the bits of a generative refinement network's binary code, avoiding heavyweight conditioning branches or source concatenation. A compact encoder converts discrete source codes into continuous coordinate-wise evidence that the backbone assimilates during refinement, and the null prompt is given editing-specific meaning (no edit, supervised by source reconstruction), producing a source-preserving state in the same representation space. A second stage diffs edited against source-preserving states to revise unresolved bits. Trained on only 0.6M pairs with under 3% conditioning parameters, the 2B model scores 4.03 on OpenVE-Bench, beating several 14B open editors, and the 8B model reaches 4.18.

HarmProfile treats harmful generations as an object of study rather than a binary attack outcome, assembling over 80,000 validated harmful artifacts from 23 frontier models across 13 families, organized into 15 harm categories and 57 subcategories. Each model gets a distributional risk profile characterized by content, severity, and variation. Models reliably produce harmful content at scale but with distinct profiles, and both harmfulness and diversity increase with capability, implying refusal rates alone understate what stronger models generate once alignment fails.

AutoResearchEval covers 100 tasks grounded in published frontier science across 7 domains and the full research lifecycle, and scores process rather than just outcomes. Eight harness-model combinations yield 800 process-annotated trajectories, distilled into a 45-pattern failure taxonomy using a human-calibrated agent-as-judge pipeline over complete trajectories and intermediate artifacts. Failures converge on a missing metacognitive loop: agents rarely check what they produced against what they found, revise, or question whether their approach was sound. The pattern recurs across all eight combinations, locating the deficit at the model level rather than the scaffold.

PACE-Bench tests whether self-evolving agents recover when execution conditions change, rather than optimizing under fixed ones. It pairs 144 source-to-target adaptations across six physics domains: a code design that works in the source fails in a mutated target with identical goal and interface, and the agent must repair it from sandbox diagnostics within a fixed attempt budget. Ten methods across four paradigms leave it far from saturated, with Reflexion plus Qwen3-14B at 35.9% overall and GPT-5.5 at 66.7% on the Statics subset. Simulator-grounded reflection beats unverified self-revision, and revealing the exact physical change does not raise the ceiling.

Saildrone announced three European construction projects: the first 52-metre Saildrone Spectre at Van Der Valk Shipyard in the Netherlands, the next-generation 20-metre Surveyor at Umoe Mandal in Norway, and rigid wing sails for the Surveyor manufactured by Southern Spars in Poland. The Spectre, unveiled in April, displaces 250 tons, reaches 30 knots, targets an anti-submarine warfare gap, supports kinetic strike and can carry two Lockheed Martin MK-70 Payload Delivery Systems. Saildrone opened its EU headquarters in Denmark earlier this year, and Danish forces began a multi-month trial of four Voyager vessels in summer 2026.

StartupBench derives agent tasks from AI products with demonstrated market adoption rather than from researcher intuition, converting real product workflows into deliverable-oriented end-to-end tasks graded with fine-grained rubrics. Under a unified harness the strongest model completes only about 30% of the suite, though partial progress is common, and complex instruction following plus domain-specific expertise dominate the failure modes. The framing is a useful counterweight to benchmarks where task selection is disconnected from actual demand.

Perplexity's Computer agent can now be triggered by emailing, forwarding a thread to, or cc'ing a dedicated address. The agent verifies the sender, reads the full thread and attachments, and executes using that sender's own connectors and permissions before replying in-thread, returning spreadsheet models, PDFs and decks as attachments. Each email task is a full session viewable on web and mobile with the same step trace and audit trail as web-app runs, and draws on the same persistent memory for team terminology and prior decisions. Computer replies only to the sender for now; reply-all for Enterprise is pending. This follows Computer in Slack and Teams.

Stability released two beta interfaces for Stable Audio 3.0. A macOS plugin in AU and VST3 formats runs as an instrument inside Logic Pro and Ableton Live, syncing generation to session BPM, producing anything from short sections to six-minute full songs, and keeping multiple takes in a playlist for comparison. The web app adds iterative natural-language editing — less low end, later build, different instrumentation — plus audio-to-audio genre and mood variation, multi-track mixing with per-track and master effects, and stem bounce. Models are trained on licensed data and users own and may distribute outputs freely.

The Information reports Anthropic is preparing a supervoting share class for its founders as it positions for a public listing, a structure that would preserve founder control of the company's direction through and after an offering. Stratechery bundles the item with Nvidia's backing of an OpenAI datacenter and Google's Spirit Airlines data purchase as part of a single week of capital-structure manoeuvres across the frontier labs. Governance mechanics of this kind matter disproportionately at a lab whose stated safety commitments depend on board and shareholder composition, though no filing details or timeline were disclosed.

Instead of curating each image-generation task in isolation, this work organizes supervision around the dependency structure between generative capabilities. Three data engines build relational supervision for text-image grounding, inter-image transformation, and image-knowledge association, caption experts align text-to-image and editing data across granularities, and a multi-stage curriculum advances task mix, concept distribution, data quality, and resolution together, with gap-aware resampling closing the loop. The pipeline produces a 440M-image T2I corpus, 120M editing pairs, and 27M image-entity pairs, used to train 3B and 6B multimodal diffusion models from scratch.

Z AI's GLM-5.3 (max) entered the Artificial Analysis Intelligence Index at 59.51, eighth of 182 models and roughly level with Kimi K3 at 59.70, against Claude Opus 5 at 63 and GPT-5.6 Sol at 61. The model is a 753B-total, 40B-active reasoning model with a 1M-token context, MIT-licensed weights on Hugging Face, priced at $1.40 per million input and $4.40 per million output with an 81.4% cache-hit discount. Per-eval scores under Index v4.1.1 include GPQA Diamond at 0.9172, Terminal-Bench v2.1 at 0.8390 and HLE at 0.4226, with an omniscience score of 14.3 against a 0.2955 hallucination rate. Output speed of 84.7 tokens per second is only marginally above the class average.

The Naval Postgraduate School and nonprofit FLEETWERX tested distributed advanced manufacturing during the 31-nation RIMPAC exercise, using an IL-4 secure-but-unclassified cloud to link more than 50 sites hosting 3D printers, CNC lathes and mills, plus robotic boats running unmanned supply legs. Some assets sat aboard participating ships; most operated from US bases, depots, universities and businesses, taking network orders and shipping finished parts commercially. Over two weeks the network received 2,741 part requests, more than the experimental system could fill — the demand signal being the actual finding rather than any single printed component.

Ground-to-air drone detection in MWIR/LWIR imagery suffers from weak thermal contrast, sensor noise, and almost no labeled data. This study tests a synthetic-first recipe: pretrain on generated infrared scenes, then fine-tune on real thermal captures. Synthetic data is enough to learn initial object representations, but even small amounts of real in-domain IR data are needed to close the deployment gap, and dataset alignment matters more than model scale. Feature-space semantic alignment is the strongest predictor of detection performance, with radiometric properties such as entropy and dynamic range contributing secondarily.

Warp Factories ships a prebuilt environment for deploying and steering cloud agents, organized around triage, specification, implementation, review and verification, any of which can be automated. Teams bring their own coding model and harness — Warp states parity between OpenAI's Codex and Anthropic's Claude Code — with integrations for Linear, Jira, Slack and Microsoft Teams. Management tooling compares performance across agent configurations, tracks total token spend, and runs self-improvement loops over the system itself. CEO Zach Lloyd said Warp currently automates 30 to 35% of its own tasks weekly, and named Stripe's internal minions system and Ramp as companies already running in-house equivalents. Access is gated behind a request page.

ChatGPT Ads expands next week to 31 European countries including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria — the largest expansion since the February 2026 US pilot, which reached eight additional markets over six months. Access initially runs through OpenAI's ads solutions team, agency partners and technology partners, with self-service Ads Manager slated for later this summer. Ads appear only on Free and Go plans. Since launch the platform has added conversion-optimized bidding beyond CPM and CPC, geo-targeting and custom audiences, and measurement via the OpenAI Pixel, a Conversions API and third-party integrations. No revenue, impression or CPM figures were disclosed.

The E-2D Advanced Hawkeye Block II upgrade passed critical design review, clearing Northrop Grumman to begin retrofits ahead of flight testing in fiscal 2029 and first delivery before the end of the decade. Block II adds an updated cockpit, increased computing power, stronger cybersecurity and an Open Mission Systems architecture that the Navy says resolves current and future parts obsolescence while enabling non-proprietary technology insertion. The aircraft carries the AN/APY-9 AESA radar for cruise-missile and aircraft detection and first flew in 2007. Northrop will also incorporate Block II into new production.

Prior Labs open-sources three pieces of relational-learning infrastructure: RelArena-α, a harness that standardizes data loading, evaluation protocols, and tuning regimes on RelBench v1; TabPFN-Rel, a relational wrapper around TabPFN-3 that currently ranks first on that leaderboard; and RPI, a model-agnostic interface for defining prediction problems over new databases. TabPFN-Rel's ranking adds to the accumulating evidence that flattening a relational database into a single table stays competitive with purpose-built relational architectures on real tasks.

RUPA moves uncertainty quantification for agents from local token-level signals to the trajectory level. An execution history becomes a directed graph whose nodes are reasoning states, tool interactions and environment feedback, connected by temporal and semantic dependency edges, and uncertainty is propagated over that graph to model how execution risk accumulates and transfers across steps. The propagated signal is fused with trajectory behavioral features and goal-alignment information into a single confidence estimate. On τ-2, Terminal-Bench-2 and GAIA with six open models, it beats token-probability and per-step confidence baselines, catching failures whose cause originates several steps before the final answer.

LIBERO-VIFO evaluates whether vision-language-action models follow visual cues they should and ignore the ones they should not. It defines eight cue families and four protocols: two testing cue understanding and authorized following, two probing unauthorized following under language-cue conflict and under empty language instructions. Across seven VLA models, cue understanding does not reliably translate into correct execution, and models will execute cue-indicated tasks with no language instruction at all, meaning an image-embedded cue can drive behavior on its own. Findings hold on scene-instantiated cues and real-robot deployment.

Feedforward transformers can only update their internal state a fixed number of times per token, bounded by depth. Recirculation adds an inference-time recurrence that lets an off-the-shelf model behave as a dynamical system and track belief state, with serial prefill but essentially no extra generation latency, and no weight training. An adaptive variant tunes only a few hyperparameters over frozen weights. On the Gemma3 family it cuts perplexity 23% across a dataset suite and lifts GSM8k accuracy 21%, positioning state tracking as an architectural fix rather than something to spend chain-of-thought tokens on.

Australia and Japan completed field trials of a co-developed high-energy laser, Project Boobook, involving Australia's Defence Science and Technology Group, Mitsubishi Electric Australia and Mitsubishi Electric Corporation. Power output and trial dates were not disclosed; the release cites more than twenty years of DSTG research and names ground vehicles and naval vessels as potential hosts. Mitsubishi Electric Australia will invest A$70 million, about US$49.8 million, in new facilities at Rydalmere in western Sydney ahead of large-scale local manufacturing. It is the first joint development between the two countries' defense industries, and both governments will also accelerate discussions on Japanese hypersonic testing at Australian ranges.

Latent 3D generators want structured tensors, but pre-optimized 3D Gaussian Splatting reconstructions are unordered, irregular, and vary in primitive count. GS-Voxel deterministically converts a 3DGS reconstruction into sparse active voxels with no per-scene fitting, keeping sub-voxel positions and rendering attributes, then encodes voxel geometry and local Gaussian attributes through a factorized VAE into sparse latents whose size scales with occupied voxels rather than a fixed primitive cap. Image-conditioned flow models trained in that space generate aerial scenes, and overlap-aware tiled inference extends synthesis past a single training crop using satellite-view conditioning.

A DQN iteratively refines grasp candidates produced by a geometric planner, operating on keypoint object representations extracted from 2D overhead images in simulation. On 300 Dex-Net objects with a UR5, the refinement converted grasps the geometric method had rejected, reported as a 100% success rate on that previously ungraspable subset, and a refined grasp transferred to a physical Delta parallel robot. The setup is narrow, top-down 2D vision trained in sim, but the residual-refinement framing is cheap to attach to existing analytic grasp pipelines rather than replacing them.

MoNe bolts a modular neural memory onto a frozen pretrained transformer to serve long contexts without retraining the backbone. It ingests context in fixed-size segments by test-time training fast-weight memory networks with layer-localized gradient updates; at query time the memory synthesizes keys and values from the query tokens alone and never re-reads context tokens. That gives O(N) preprocessing with O(1) per-query cost and peak GPU memory independent of context length. At 128K tokens it cuts compute and peak memory roughly 80% versus in-context learning for 6.4% parameter overhead, and holds up on RULER needle-in-a-haystack and word extraction past the backbone's native window.

Memory-based self-improving agents are typically reported from single runs on a fixed task stream. Re-evaluating two such methods across multiple seeds and randomly shuffled task orders exposes two problems: evaluation noise in multi-step environments is already high and the self-improvement loop amplifies it, and reported gains depend heavily on task order because default orderings act as an implicit curriculum. Inspecting the memory banks points to task and environment underspecification as a driver; adding detailed rubrics and environment feedback to memory construction recovers part of the degradation but leaves clear gaps, implying other unidentified factors.

A red-team measurement of indirect prompt injection against DeepSeek Harness, run through AI-Infra-Guard: 14,560 controlled executions spanning 16 indirect-content channels, text and file carrier modes, 35 payload objectives and 12 attack methods, with the agent loop, tool registry and model adapter left intact and sensitive sinks replaced by local fixtures so attempted actions are logged without side effects. Peak attack success reaches 25.5% for hidden Unicode in file mode, 17.0% for fake-completion in text mode and 16.0% through the skills channel. A deterministic rule judge and an LLM judge disagree notably on partial compliance (7.3% versus 2.0%), and the code is public.

Tokenizers are usually chosen with little more than fertility and compression rate as evidence. TokEval adds metrics for linguistically and structurally meaningful properties, including UTF-8 character boundary integrity and digit place-value boundary alignment. To check whether these predict anything, the authors run controlled pretraining sweeps varying only tokenizer training mixture, pretokenization strategy, and training algorithm, then evaluate on bits-per-byte plus linguistic, math, and code benchmarks. Information-theoretic metrics track language modeling ability with Spearman rho up to 0.80, while structure-sensitive metrics such as digit and line-break handling correlate with task accuracy, suggesting intrinsic measurement can substitute for some pretraining sweeps.

Per-field accept-or-review with a selective risk budget is the contract document-extraction systems are sold on, and this work shows the obvious procedure silently breaks it. On 13,859 extracted fields from 800 CORD receipts (49.0% correct), three failure modes appear: document-level clustering with design effects of 1.84 to 2.45, score-refit leakage that violates alpha of 0.10 in 95% of splits, and a tie-mass pathology that collapses the threshold grid. A fit/val split restores expected-risk control at coverage 0.318 and risk 0.096 but still overshoots alpha in 47.5% of resplits, while Mondrian Learn-then-Test gives per-group PAC certificates that are nearly vacuous at the document-iid tier.

aDSL co-designs a domain-specific language and the multi-agent system that writes it, on the premise that LLMs reason better over semantic structure and spatial relations than over brittle absolute coordinates. Geometry is manipulated through composable relational operators, and a training-free plan-execute-critic loop synthesizes code and repairs constraint violations using execution feedback. It outperforms prior LLM-based baselines on text-to-shape and image-to-shape while keeping programs explicit and editable, and extends to articulated objects and structured scene composition.

Self-evolving agents that distill experience into skills or memory are usually scored on post-evolution accuracy alone. Auditing SkillOpt, Agent Workflow Memory, and ReasoningBank in a simulated e-banking environment with sealed evaluation endpoints and independent state replay shows capability and attack surface rising together. On Qwen 3.7 Flash, SkillOpt lifts benign utility from 0.741 to 0.837 while exposure to injected content climbs from 0.820 to 0.943, overall attack success rises from 0.496 to 0.530, and unauthorized financial state changes reach 0.685. ReasoningBank improves utility without raising aggregate ASR. AWM exposed a separate artifact-executor mismatch that alone swung utility from 0.319 to 0.756.

A benchmarking study of off-the-shelf detectors for automatic target detection and recognition, run on a newly collected military-relevant dataset with heavy occlusion and small targets. Six YOLO generations and two DETR variants are each evaluated out of the box and after fine-tuning on VisDrone as an air-to-ground proxy domain. Larger models win, DETR variants compare favourably to YOLO, and out-of-domain fine-tuning gives modest air-to-ground and small-object gains, but every model still fails on small aerial targets, so in-domain data remains the binding constraint.

Most diffusion acceleration work targets solvers and samplers; the schedule of sampling timesteps itself is usually taken from defaults or from theoretical surrogates for sample quality. OYS instead treats timestep selection as black-box optimization and tunes the target metric directly with Bayesian optimization. It beats both default schedules and Align Your Steps on text-to-image and improves inpainting and other image tasks on quantitative metrics and human evaluation. The method needs no training, applies to distilled models, and works with Euler and DPM-Solver++ alike; a 5-step OYS schedule retains 89% to 94% of 50-step quality at a tenth of the inference cost.

Linguistic olympiad puzzles invert the usual reasoning setup: the solver must infer the rule system before reasoning inside it. This open competition used unseen problems from the 2026 International Linguistics Olympiad individual contest, graded both automatically and by the official jury under the rubrics applied to human contestants. It drew 731 submissions from 46 teams under a hard budget of one T4 and 30 minutes, plus 15 unconstrained models, with Claude Opus 4.8 earning a gold-medal-equivalent jury score while the constrained entries landed in the bottom 5% of contestants. Scale was not decisive: 14B submissions beat models twice their size, and automatic metrics preserved the jury ranking but inflated weak systems by about 13 points.

Teams migrate off deprecated model versions on the strength of aggregate benchmark deltas, and this study measures what that averaging hides. Querying 900 public benchmark items 50 times per model across three upgrades in one commercial product line, then classifying each item as improved, regressed, equivalent, or inconclusive under false-discovery-rate control with a practical-significance threshold, it finds reliable gains and reliable regressions coexisting in all nine migration-benchmark cells. Edges with aggregate gains up to 7.3 points still contain up to 8.3% reliably regressed items. Response-level archives are released.

Automated security patch backporting tools report above 80% success, but usually on single repositories or narrow version ranges. Porting Benchmark supplies 1,234 cases spanning cross-version, cross-branch, and cross-repository settings under one evaluation framework, and re-scores five tools covering program analysis, LLM prompting, and LLM agents. Aligned evaluation reorders the field: PortGPT and TSBPort hold up while FixMorph and Mystique degrade sharply, and the best commit-level success rate falls from 85.2% on simple Type-I patches to 24.0% on Type-IV. A 45-case dynamically validated subset with proof-of-concept exploits shows reference matching under-credits hard adaptations while hiding real integration failures.

An LLM-as-judge study argues grading reliability comes from the rubric, not the judge. A frontier model reads source documents once to extract questions and rubrics; six cheap configurations across two families and three reasoning-effort levels then produce 3,456 per-question grades on 24 open-ended answers. Answer identity explains 95.6% of score variance against 0.2% for judge identity, and raising judge reasoning effort shifts scores by at most 0.006 of full marks versus 0.143 for writer effort. Ablations isolate the cause: stripping criteria and levels changes nothing, but removing the official answer collapses ICC from 0.888 to 0.628.

InnerExpert exploits routing-level signals that only exist in sparse architectures, such as router entropy, expert disagreement and expert usage patterns, and combines them with standard transformer activations into compact per-token feature vectors. A lightweight classifier is trained on labels from an LLM-as-a-judge pipeline, which removes the manual annotation bottleneck and allows retraining as the base model changes. Across five datasets and two MoE models it reaches up to 0.91 answer-level and 0.76 token-level AUROC using a single forward pass. The token-level resolution is the practical part, since it localizes hallucinated spans for targeted intervention rather than flagging whole answers.

Three MoE models with 3.6-4.0B active parameters are fine-tuned to reason in Greek, and accuracy barely moves; the authors argue the benchmark cannot resolve the difference anyway, since reseeding alone shifts scores by 7.7 points. The behavioural changes are what matter. Base models produce Greek reasoning traces in 0 of 1,000 cases; after SFT roughly 98% of traces match the question language with general ability essentially preserved. RLVR then fixes what SFT leaves broken, cutting format fallback from 24% to 2.5% and reasoning-channel answer leakage from 3.5% to zero.

DAS generates publication-oriented literature surveys by separating reusable paper analysis from topic-specific manuscript construction, backed by DAS-2M, a continuously updated metadata lake of survey-oriented representations for about two million papers. Its agents maintain explicit literature, organization, writing and finalization states, using candidate-grounded taxonomy planning, reverse paper-to-section routing and hierarchical claim and citation planning, with semantic review reactivating only the affected writing states so repair stays a scoped closed loop with deterministic validation. On the accompanying 30-topic DAS-Bench scored over 16 criteria it leads every dimension with an overall 4.34 against 4.03 for the strongest competitor, and blinded experts prefer it to naive RAG on 27 of 30 topics.

OpenAI and CodeAI announced a partnership on high-school AI literacy timed to the ChatGPT for Teens launch, citing that only 16% of high school leaders say all their students learn the technical knowledge to understand AI in class, against 75% of students in a CodeAI survey saying AI understanding will matter more in future. Five workstreams over the next year: a joint advisory council on child development and learning science that will also shape ChatGPT for Teens, an Hour of AI, a first Builders Challenge with OpenAI mentorship, expert consultation on CodeAI's free year-long AI Foundations course, and Career Journeys sessions. No funding amount or reach target beyond millions of students was disclosed.

A position paper arguing that current AI co-scientists are researcher-agnostic: given a goal they optimize novelty, validity or predicted reviewer score while ignoring who will actually use the output, even though what counts as novel or feasible depends on a researcher's prior work, methods and collaborators. The proposed framework threads a graph-grounded researcher representation through retrieval, hypothesis search, experimentation, writing and review, and calls for evaluation grounded in the individual rather than generic quality scores. The named failure mode is that distinct researchers issuing the same goal receive essentially identical research. No system results are reported.

Concept erasure in text-to-image diffusion is normally probed through text, testing whether a prompt can still summon the erased concept, which leaves open whether the visual knowledge survives. TINA+ probes it visually via optimization-based diffusion inversion under a null-text condition, avoiding the text pathway. The authors also flag a confound: unconstrained inversion finds spurious trajectories that let even a randomly initialized model reconstruct the target, so they add trajectory regularization penalizing paths below the expected marginal energy evolution. Across twelve erasure methods and four tasks, current methods mostly sever text-image links rather than remove the visual knowledge.

rl-triton recasts seven RL credit-assignment estimators (GAE, V-Trace, Retrace(λ), TD(λ) returns, discounted returns, eligibility traces and episodic prefix sums) as instances of a single first-order linear recurrence solved by one associative scan in O(log T) parallel steps, with algorithm-specific fused Triton kernels building the recurrence coefficients on-chip. Terminated and truncated episodes are handled explicitly and the scan operator is verified algebraically. In the massively parallel simulation regime of thousands of environments and short rollouts, full-call speedups run 1.6x to 5.7x over a vectorized torch.compile baseline, widening at longer sequences because the baseline pays an HBM round trip per scan stage. Released open source.

A systematic benchmark of four pretrained molecular language models as encoders across six virtual libraries covering drug discovery, organic materials and catalysis. Native embeddings vary substantially in discovery performance from library to library, while classic molecular fingerprints hold up as a consistently strong baseline, which is an awkward result for foundation-model-first pipelines. Fine-tuning the encoders on structures drawn from the target library consistently improves sample efficiency, and several domain-adapted encoders become the top representation on their benchmark tasks. The takeaway for virtual screening and self-driving lab loops is that adaptation to the search space matters more than the choice of pretrained model.

RLVR wastes rollouts when every prompt gets the same exploration budget, but online difficulty estimation usually requires dedicated probing or suffers cold start and stale feedback. This work builds a similarity graph over training samples, couples neighbours through a Potts prior over latent difficulty states, aggregates rollout outcomes per state with a Beta-Binomial model, and updates assignments online by mean-field variational inference. The estimator drops into existing sample-selection and rollout-allocation schedulers and improves results across several base models and benchmarks without extra generation.

Parametric CAD constraints give a controlled setting for asking whether decodable structure in a frozen model is actually used. Probing six decoder-only LLMs across four properties, the authors find pretraining clearly improves linear decoding of local pairwise geometric relations, but sketch-level degrees-of-freedom status is already highly decodable from randomly initialized weights and barely improves with pretraining. More pointedly, decodability does not imply control: generation often fails to express the probed information, activation-restoration effects vanish with depth while probe accuracy persists, and mean-difference steering does not reliably move outputs.

Undersampled MRI reconstruction uses complex-valued k-space, while sparse phase retrieval suggests magnitude-only measurements carry complementary information that has gone unused for lack of a way to acquire it for free. The authors show k-space magnitudes are strongly consistent across time-frames in steady-state dynamic acquisition, then build C+Mag, an ADMM-unrolled physics-driven reconstruction with a magnitude-aware data-fidelity term, using quadratic smoothing and momentum updates to handle its non-differentiability and non-convexity. On retrospectively undersampled cine and phase-contrast flow MRI plus prospectively undersampled real-time cine, it improves artifact suppression and phase preservation over standard unrolled baselines, confirmed by blinded reader evaluation.

Language-model log anomaly detectors score well on detection but are badly calibrated, assigning high confidence to wrong predictions, especially on anomalous logs under heavy class imbalance. The authors show this persists even when standard calibration metrics look healthy, which is the failure mode that matters for automated operational monitoring. LoRD is a lightweight post-hoc fix: it fits prediction-route-specific reliability models on latent representations of correctly classified validation samples, estimates reliability via route-wise reconstruction distance, and selectively recalibrates only high-risk predictions. Across four large log benchmarks and several detectors it suppresses overconfident errors without hurting detection.

CADP reframes academic-document parsing as code generation: reconstruct a page as contextual LaTeX plus executable Python, so tables, formulas, charts, and pseudocode keep their structure, data, and logic instead of being flattened into Markdown. Because the output compiles, reconstructions can be re-rendered and verified directly against the source page. CADP-Bench supplies expert-verified full pages with a re-injection compilation protocol; current frontier MLLMs and an exploratory multi-agent baseline both fall well short of high-fidelity executable reconstruction.

The Travelling Thief Problem with Drone couples item selection, ground-vehicle routing, and drone launch and rendezvous timing, where accumulating payload slows the vehicle and every pickup shifts the synchronization schedule. The authors give a MILP that solves small instances exactly, plus metaheuristics and an attention-based deep RL policy for larger ones. A hybrid where the learned policy seeds a short annealing run recovers most of the metaheuristic's solution quality at a fraction of the compute, though the largest instances still need the full baseline budget. Sensitivity analysis puts the rental ratio, not fleet parameters, as the dominant profit driver.

Because LLMs emit executable code, evolutionary search can operate over procedural content generators rather than individual levels. Runs evolve complete Python generators through language-model mutation and crossover in Sokoban, Zelda, Dangerous Dave, and Lode Runner. Continual Abstraction Discovery (CAD) extracts reusable primitives from high-fitness programs into a run-specific helper module that later generations import. A 2x2 design crossing CAD with access to a hand-written domain API over 160 runs of 50 generations shows CAD raises mean final best fitness in all eight domain-by-API comparisons, and the learned libraries repeatedly rediscover validation, reachability, and structural utilities.

Reflex-Guard replaces judge-model and cloud moderation guardrails with local classification: jailbreak-aware preprocessing, compact sentence-transformer embeddings, and seven fast binary classifiers, which also removes the need to route user prompts to an external endpoint. On a balanced 30,568-sample evaluation it reports 95.9% recall on harmful prompts at 37.6 ms end-to-end, against 255 ms for Llama Guard 2 and 723 ms for SafeDecoding, and catches all GCG suffix and Base64-encoded attacks at the default threshold. DrAttack-style structured prompts need the threshold dropped to 0.03.

A controlled LoRA-GRPO study on RWKU compares four reward designs for LLM unlearning: lexical suppression, anti-refusal shaping, rubric-scored broad answering, and an explicit refusal contrast, each with and without SFT warm-up. The central finding is that reward optimization and behavioural unlearning come apart. Forget scores, held-out completion audits, terminal rollout audits, and training dynamics can each support a different conclusion, with the disagreements traced to reward-hacking endpoints, policy-support limits in GRPO, and probes that miss endpoint changes.

BEAR-Bench targets reasoning over text-dense business and scientific documents rather than information extraction, with 1,000 human-annotated questions in English and Russian, a pair chosen because existing multimodal benchmarks skew English or Chinese and are rarely self-contained. Sixteen proprietary and open-weight MLLMs, including recent frontier systems, leave clear headroom. The collected model outputs are then reused as a testbed for comparing hallucination-detection methods, measuring not only how often models fail but how reliably those failures can be flagged.

Bias benchmarks force models into a fixed answer format, and survey science has long known that format shapes answers. Testing three instruction-tuned models on BBQ and OpinionQA under closed-ended, Likert-scaled and open-ended formats with everything else held constant, measured bias changes enough to reverse model rankings. The mechanism is that each format elicits a different response behavior: forced-choice selection, scale-based distributions, and refusal in free-text generation. The practical implication is that reported bias numbers and human-alignment scores are format artifacts unless benchmarks adopt multi-format designs.

Hybrid systems that use an LLM planner over an RL controller usually leave the status of the LLM-derived reward signal informal. This paper casts the setup as a goal-augmented MDP and shows that if the LLM's per-state progress score is used as a bounded potential function, the resulting shaping term preserves the optimal policy set even when the scores are wrong, a stronger guarantee than generic LLM-as-reward schemes offer. Numerical checks on a small MDP cover four potential configurations, including an adversarial one scaled to twenty times the base reward magnitude.

Model-based RL agents learn latent world models whose representations are entangled with the training task's reward, which limits reuse. This work restricts reward prediction to a structured symbolic subset of the latent state while leaving observation reconstruction to the neural component. Decoupling the two allows the same world model to be applied zero-shot to any new reward function defined over that symbolic space, with no further environment interaction, and the authors report stronger generalization than purely neural baselines.

TSN4PI is a two-module pipeline for tracking ideological signals in social media text over time. The detection module (PIDN) uses LLMs with style transfer and unsupervised domain adaptation to classify posts and filter out off-topic content in noisy cross-domain data, addressing label scarcity and annotator cost. The prediction module (PIPN) applies temporal graph neural networks to forecast future shifts, and the framework reports presence, intensity, and change over time. The authors release two large-scale datasets for noncommercial research and run case studies on X and Truth Social data.

A monograph proposing a typed framework for reasoning about when elimination or compression destroys distinctions that a downstream task actually needs. It separates local solvability, global realizability and finite-sample certifiability, and tries to distinguish genuine architectural obstruction from approximation error, generalization gap and implementation bugs, with worked interfaces for integrability, representation admissibility, resource limits and deployment overlap. Applications span sparse model selection, distribution-free prediction, observational treatment policies, routed expert and retrieval systems, and learned score fields. The operational claim is that a persistent performance floor may only be blamed on architecture once the deployment contract, budgets and validation rule are fixed in advance.

Human-designed assessments implicitly assume models and people draw on comparable underlying constructs. This study runs exploratory factor analysis separately on human and six-model response sets from quantitative reasoning and chemistry assessments, then has subject-matter experts blindly interpret the resulting factor graphs. Experts could name most human-derived factors but could not ascribe meaning to any model-derived factor in quantitative reasoning and only half of them in chemistry. The result is evidence that the latent structure governing model performance on these instruments is statistically opaque and distinct from human reasoning, which weakens the case for reusing human psychometric assessments as model evaluations.

AVShift is a German authorship verification benchmark of over 150,000 text pairs spanning three genres and 21 years, built so that cross-genre, temporal and AI-era shifts can be varied within one framework instead of studied in isolation on English data. Benchmarking feature-based, embedding-based and LLM-based verifiers, fine-tuned LLMs generalize best across genres and benefit substantially from stylistically diverse training data. Temporal drift emerges as one of the strongest degradation factors, with accuracy falling as the gap between documents grows, while no measurable AI-era distribution shift appears in this corpus. Data and code are released.

Rather than predicting unsteady aerodynamic lift directly, an LSTM is trained to learn the residual between high-fidelity CFD lift and an analytic Wagner-function baseline for the NLR 7301 airfoil under prescribed pitch and plunge motions in transonic flow with shock motion. Against a direct predictor, the residual model performs best when its inputs align with the Wagner formulation variables, giving lower error, more consistent behavior across training runs, and smaller degradation in leave-one-out and leave-family-out tests, though direct prediction still wins on some high-frequency cases. The pattern supports residual learning as a modular way to augment classical low-order aerodynamic theory.

CoAL-RAG routes legal queries to different retrieval strategies based on estimated complexity, avoiding both over-reasoning on simple questions and opaque answers on hard ones. Reasoning demand is quantified from the logical structure of the question, and the disagreement between semantic and keyword retrieval serves as an indirect complexity signal that selects the strategy and filters the retrieved context. Reported gains include a 42.5% BLEU improvement on Chinese legal benchmarks and ROUGE-L at 3.6 times a knowledge-graph baseline, with accuracy holding up on LexGLUE and CaseHold as a cross-jurisdiction check. The comparisons are against relatively modest baselines rather than strong long-context systems.

Mapping invoices to the correct general ledger code is an accounting judgement depending on the purchasing business, the vendor and the invoice text. Analyzing SBERT and DeBERTa embeddings of a financial corpus, the sentence-embedding space is globally anisotropic but built from locally isotropic clusters that correlate strongly with vendor identity, extending earlier token-level findings to sentence embeddings. Fine-tuned on a single GPU, SBERT reaches 0.96 classification accuracy, above both a zero-shot LLM and a vendor-identity baseline, and hits 0.9 F1 on a new client with roughly 100 client-specific invoices. Structuring the input in a way that helps a human reader did not help the model.