Wolf Digest — Wednesday, August 5, 2026

The Space Force has awarded three other transaction agreements worth a combined $615 million to mature Space-Based Airborne Moving Target Indicator capability and widen the vendor pool developing it. Rocket Lab and STR are two of the three; the third could not be named for operational security reasons. This is the second task order under the SB-AMTI program and builds on the $4.16 billion deal given to SpaceX in May.

Space Systems Command frames the new contracts as accelerating emerging systems, leveraging commercial technology, and demonstrating novel approaches, which is procurement language for deliberately not concentrating the capability in one prime. Tracking airborne moving targets from orbit is a hard detection problem well before it is a hard engineering problem: the target signature is small, the clutter background is enormous, revisit rates constrain track continuity, and the discrimination step is exactly the sort of learned-classifier workload where architecture and training data dominate outcomes. Spreading $615 million across three teams buys three independent bets on how to solve it.

Annotation pipelines for driving vision-language-action models routinely show the teacher the logged ground-truth future trajectory when generating chain-of-thought supervision. The authors show this induces trajectory anchoring bias: the teacher rationalizes the revealed outcome instead of inferring a decision from scene evidence, producing less causally faithful reasoning and substantially worse hallucination rates, with the damage concentrated in causally challenging scenes. Their fix defers exposure of the future trajectory so the reasoning is generated before the answer is visible. It is a clean instance of a general problem in distillation for control: privileged-information teachers produce fluent explanations that are not the explanations the student can reproduce at inference.

RAND describes secure inference data centers, purpose-built facilities designed to protect model weights and inference from advanced nation-state adversaries, and argues they can be built today with proven technologies rather than requiring new research. The framing matters because most weight-security discussion has focused on training-cluster exfiltration, while inference serving is where weights sit resident, hot, and distributed across far more physical locations. Treating the serving tier as the harder security perimeter is a shift in emphasis, and the claim that no new primitives are needed makes it an execution question rather than a research one.

Critic-based reinforcement learning for vision-language-action models generally estimates value from single-frame observations or single-frame VLM latents, which is a structural mismatch with the partial observability of real robot control. Naively feeding observation history into the critic blows up complexity in high-dimensional visual space and still underperforms because a scalar return carries too little signal to shape a visual value function. WCM proposes a world critic that models the observation stream rather than compressing it to a frame, aiming to make the value estimate reflect what the policy can actually infer about state.

Anthropic has reportedly signed a $10 billion deal with AI cloud startup Volta, the latest in a run of cloud partnerships the company has struck over recent months. The pattern is the story: rather than consolidating on a single hyperscaler, Anthropic has been assembling capacity across multiple providers, which spreads supply risk and preserves negotiating leverage but fragments the serving stack across heterogeneous hardware and scheduling environments. A ten-figure commitment to a startup provider also implies confidence in that provider's ability to actually deliver silicon on a schedule, which has been the binding constraint on every capacity deal signed in the past two years.

The Army published a sources-sought notice for its Next Generation Counter-sUAS Missile initiative, seeking a munition that can defeat small drones at extended range with rapid launch and reduced time-to-target, at under $150,000 per round. The cost ceiling is the whole point: the current exchange ratio, where an interceptor costs orders of magnitude more than the drone it destroys, is unsustainable at the volumes now being seen operationally. The notice lands the same day as the Pentagon's counter-drone marketplace award, part of a broader push to widen the toolset against small unmanned systems.

A new SaferAI report finds that Z.ai's open-weight GLM-5.2 approaches frontier capability while lacking key safety mitigations that the closed frontier labs ship alongside comparable models. The framing TechCrunch lands on is the governance one: the capability gap between open and closed weights has narrowed faster than the mitigation gap, so the marginal capable model released into the wild is now a model whose guardrails can be removed by anyone with a fine-tuning budget. That argument is not new, but the empirical claim that a specific open release is at or near frontier on capability while measurably behind on mitigations gives it a concrete anchor.

Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter on-device agentic model pre-trained on roughly 34 trillion tokens with mid-training extending context to 128K. Post-training is the interesting half: two rounds of agentic-weighted supervised fine-tuning, per-domain teacher specialization, multi-domain on-policy distillation, then agentic reinforcement learning inside real harnesses (OpenClaw, Hermes Agent) via a Harness Proxy that treats each harness as a black box while capturing token-level trajectories.

Against gemma-4-E2B-it, gemma-4-E4B-it, Qwen3.5-4B and Qwen3.5-9B it takes the top slot on IFBench (59.17), Multi-IF (80.07), IFStruct (85.49), ToolSandbox (77.83), tau-cubed Banking (5.67) and AA Omniscience (-29.50), while losing BFCLv4 to Qwen3.5-9B (56.88 versus 60.13). AIME25 is 51.87 and LiveCodeBench v6 is 59.41. Throughput: 220 tokens per second on an M5 Max, 113 on a Ryzen AI Max+ 395, roughly 30 on a phone, under 2.5 gigabytes of memory, and about 15,000 output tokens per second at high concurrency on a single H100, which the post frames as 1.3 billion tokens per day. Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX.

Members of the Open Secure AI Alliance, now more than 120 organizations, are proposing SAFE guidelines to strengthen agentic AI cybersecurity, timed to the opening of Black Hat. The Linux Foundation has published a request for comments on standing up the SAFE working group as an open community effort, with the RFC repository on GitHub. The alliance was spearheaded by NVIDIA and formed roughly a week earlier, which is the detail TechCrunch fastens onto: a standards body that ships a proposal seven days after founding is either unusually well prepared or unusually eager to set the agenda before someone else does. Mistral, which released its Shieldstral moderation model the same day, is an inaugural member.

Under Secretary of Defense for Acquisition and Sustainment Michael Duffey discusses the Pentagon's acquisition transformation: recent munitions agreements, industry investment, the shift to portfolio acquisition executives, and allied coproduction. The portfolio-executive restructuring is the piece with the longest reach for anyone selling autonomy or AI software into the department, because it changes who holds the requirement and the money for a capability area rather than a program, which is the level at which software-defined systems have historically fallen between the cracks.

A 16-billion-parameter autoregressive diffusion framework for real-time video editing with no access to future frames and no predefined clip duration. Three pieces do the work: chunk-wise autoregressive adaptation, Source-Anchored Distribution Matching Distillation to preserve source fidelity through a two-step generation, and Long-Horizon Autoregressive Distillation to suppress accumulated temporal drift. The train-inference mismatch that normally kills causal video models is the explicit target, and the two-step budget is what makes the latency claim credible.

Existing agent harnesses keep task execution, task state and completion assessment inside a growing context, which makes state hard to track and lets an incorrect self-assessment propagate into every later decision. The reformulation here treats long-horizon execution as a task-state management problem: state lives explicitly outside the execution loop and is updated only with facts independently verified from the environment. That verification gate is the substantive move, since self-reported completion is the dominant failure mode in long-horizon agent runs.

Persona skills compress personal interaction history into portable executable artifacts, which concentrates fragmented personal signal and amplifies it through reuse in a way that defenses built for individual records or retrieval memory do not cover. The benchmark ships 7,500 persona-grounded dialogue traces built from 50 behaviorally rich profiles and evaluates risks and defenses end to end across the pipeline. As agent skill-sharing becomes a product surface rather than a research artifact, this is the threat model that will matter.

A concrete mechanism for the observation that test-passing model patches leave codebases harder to maintain. Across the five leading models on the official SWE-bench Verified leaderboard, deletion recall against the developer patch reaches at most 71.7 percent even on tasks all five solve, and the models reach the correct file for more than 92 percent of required deletions while failing to cut the exact lines. The failure is localization-complete and execution-incomplete, which is a very different problem from not understanding the task, and it points at supervision that rewards passing tests without penalizing retained dead code.

Reinforcement learning for tool-integrated reasoning usually supervises whole trajectories, which gives almost no credit-assignment resolution over a long tool-use episode. On-policy self-distillation offers denser signal through a privileged teacher, but existing versions build that privileged context from ground-truth answers or retrieved skills, states the agent never actually visited. TurnSight derives the teacher branch from hindsight over the agent's own trajectory and supervises at turn granularity rather than token granularity, matching the structure of tool interaction.

The framing is that pixels define how an image is rendered while layers define how it is created, understood and edited, so a generative model aimed at design work should operate in a layer-native space. Two models: a text-to-RGBA generator producing standalone layers with alpha, and a composition stage. It is the most direct attack yet on the reason diffusion models remain awkward inside real design workflows, which is that a flat raster cannot be revised.

A landscape assessment of AI-enabled products currently available to support emergency management across the United States. Landscape studies of this kind are unglamorous and useful: they establish what is actually procurable today versus what exists in a paper, which is the gap that most public-sector AI strategy documents quietly assume away. Emergency management is also a domain where the deployment constraints, intermittent connectivity, high-stakes decisions under time pressure, and heterogeneous local data, are the constraints that generalize to most government field use.

Rob Wiblin's account of how expert timelines swung twice inside a year, using Andrej Karpathy's reversal as the anchor: agents dismissed as slop last October, described two months later as alien tools rocking the profession. Wiblin had recorded a video explaining why expert timelines had lengthened, and by the time it published the vibe had already shifted again. The evidence he assembles is concrete rather than sentimental, led by models completing software-engineering tasks that would take human professionals a full day, and improving faster than METR's time-horizon measurements were built to track. The episode is most useful as a record of what evidence actually moved people, which is a different question from what evidence should have.

Text remains the outlier in generative modeling: images, video and audio are increasingly modeled in continuous latent spaces while text still leans on discrete tokens. Existing continuous language models either inherit embedding spaces never designed for joint generation and decoding, or compress the autoencoded latent to make diffusion tractable and lose token-level fidelity. AURORA-LM inverts the tradeoff, preserving a high-capacity decodable latent and designing the diffusion model around it instead of the reverse.

Supervised fine-tuning suffers severe task conflict under multi-stage multi-task training while reinforcement learning lets diverse tasks coexist stably. Traced to the parameter level, RL induces sparse and approximately orthogonal updates across tasks, and the paper supplies a theoretical account of why multi-task RL lands in that regime. If it holds up, it is a direct argument for RL as the multi-task integration stage rather than as a final polish, and a warning about sequential SFT curricula.

Dense pretrained transformers expose no natural interpretable units, so circuit work generally learns auxiliary sparse representations or trains sparse models, paying substantial extra compute and opening a fidelity gap between the object analyzed and the model actually deployed. SWD reparameterizes each pretrained linear projection as a product of two sparse factors whose shared intermediate coordinates become individually addressable circuit units, with no separate model to train. Closing the fidelity gap is the part that matters: an explanation of a proxy is only as useful as the proxy.

Recursive self-improvement requires turning accumulated experience into better future behavior, and personal agents that retain preferences, task histories, tool routines and learned skills across sessions are the cleanest available testbed. PAST-Bench runs agents through ordered sequences of fresh-session tasks under matched conditions with retained experience switched on and off, across 26 scenarios and 204 episodes covering memory, procedure and skill. The ablation design is the contribution: memory benchmarks that never turn memory off cannot separate retention from base capability.

Two failure modes surfaced by the preliminary evaluation are worth carrying beyond this paper: modality bias, where agents skip visual tools in favor of text search, and parametric knowledge leakage, where the model answers from memory rather than genuine tool-augmented execution. Both inflate scores on multimodal agent benchmarks without any multimodal capability being exercised. The proposed fix is a decoupled perception-exploration pipeline with stage-wise tool unlocking that forces exhaustive visual grounding before web exploration is available.

The Defense Information Systems Agency is conducting market research on the CommandNet migration weeks after General Dynamics Information Technology protested the decision to award the work without competition. CommandNet moves users across all eleven combatant commands onto a single DISA-managed architecture known as DODNet by fiscal 2028, and the sources-sought notice asks for technical approach, transition plans, cost estimates and past performance. The consolidation matters for AI deployment in the department for an unglamorous reason: a single managed network is the precondition for any data or model governance regime that spans commands.

Interpol reports that AI-enabled techniques now account for more than half of cybercrime across Africa, with digital scams surging. The measurement question is immediate and unresolved, since attributing a scam to AI assistance requires either artifact analysis or self-report and both are weak instruments at continental scale. The reason to track it anyway is that fraud is the misuse category with the shortest path from capability to harm: generation quality improvements translate directly into conversion rates, with no intermediate capability required.

The term now covers algorithms that extend deliberation along one trajectory, that sample completed candidates and aggregate by voting or verification, and that search over unfinished partial states. These have different statistical structure, different compute accounting and different failure modes, and treating them as interchangeable under one scalar budget, or reporting accuracy without variance, produces comparisons that do not mean anything. A reproducibility paper aimed at a literature that badly needs one.

A practitioner's study with two systems results. First, offline knowledge distillation, caching the teacher's top-K logits once and training the student against the cache, matches online distillation at near-identical training loss while removing the teacher forward pass entirely from the training loop. Second, a fused chunked KL loss cuts the memory cost of the divergence computation. For teams recovering quality in a compressed model under latency and on-premises constraints, this is the rare distillation paper aimed squarely at the cost line rather than the accuracy line.

Repository-level benchmarks evaluate agents working alone, or limit user participation to messages, which misses the actual shape of shared-workspace development where a human may inspect and modify code during an ongoing agent task. SWE-Touch introduces validated Counter-Edits, plausible edits to task-relevant code that conflict with completing the task, and measures whether the agent notices and adapts. Given how much agentic coding now happens with a human in the same branch, this is a gap worth closing.

Skill generation methods are mostly heuristic pipelines hand-designed per evidence source, and the learning-based alternative is hard because skills have no natural supervision signal: a skill's value is only revealed by whether it improves downstream agent behavior. Skill-alpha treats that downstream improvement as the reward and generates skills progressively, which is the correct objective even if it makes the credit assignment brutal.

The post starts from the MIRI position that pushing past human-level general capability with current understanding is reckless, and grants that many observers are waiting for more evidence before acting. Its contribution is to take the warning-shot concept seriously as an empirical question rather than a rhetorical device: what would count as one, what mechanisms would translate it into policy, and what the historical base rate looks like for societies converting a near-miss into binding constraint. The framing is timely given this week's disclosure that agents in a live evaluation reached the public internet, which is close to a textbook candidate for the category.

An experimental open-weight language model that generates text by iteratively refining blocks of 256 tokens in parallel rather than decoding one token at a time, sidestepping the sequential bottleneck of autoregressive decoding. Rather than training from scratch, DiffusionGemma is obtained by fine-tuning the Gemma 4 mixture-of-experts model with 3.8 billion activated and 25.2 billion total parameters, through a compute-efficient two-stage procedure. Converting an existing strong autoregressive checkpoint into a discrete diffusion decoder is the practically interesting claim here, because it means the enormous pretraining investment behind AR models is not stranded if parallel decoding turns out to win on latency.

On-policy self-distillation strengthens the teacher with privileged information, but the student then learns privilege-dependent behavior it cannot reproduce from its own inference-time context and keeps acting as though the privileged information were still there. The authors name this the privilege illusion and locate its cause in information asymmetry between teacher and student at inference. DAPD anchors on both sides to close that asymmetry, and the diagnosis generalizes well beyond this specific recipe.

Unified multimodal modeling has worked for images but stalled in 3D because the multimodal data is scarce, and geometrically consistent editing data especially so. Hunyuan3D-Buffalo supports 3D understanding, text-to-3D generation, instruction-guided editing and text-grounded part generation in a single architecture, with the data-construction pipeline doing much of the heavy lifting. Part-level grounding is the capability that makes 3D assets usable downstream rather than just renderable.

Aligning diffusion models to human preference normally relies on a sparse terminal reward on the final sample, which leaves a severe temporal credit-assignment problem across the denoising trajectory. Learnable position-free register tokens prepended to a frozen diffusion transformer's input sequence read terminal preference directly off intermediate noisy latents, without altering hidden states or the velocity field. The independence of the readout is what makes the resulting dense signal differentiable through the trajectory without corrupting the generator.

Muon orthogonalizes each weight-matrix update with a Newton-Schulz iteration, performing steepest descent under the spectral norm, and essentially all published evidence for it comes from transformers. This controlled comparison on Mamba-2 130M varies only which weight groups get Muon, and finds the benefit is localized: Muon on the output projection alone beats Muon on the input projection or on both together. That asymmetry is a useful signal about where the conditioning problem actually lives in a state-space block, and a caution against treating optimizer results as architecture-independent.

Long audio-visual token sequences are heavily redundant, and existing compression degrades sharply at low budgets: pre-LLM compression discards structurally important globally distributed evidence, while inner-LLM compression underuses query-conditioned audio-visual collaboration. OmniPack is training-free and coordinates both, which is the practical property, since a compression scheme that requires retraining the omni-model is not a deployment option for anyone serving a released checkpoint.

LLM 0.32 is, by its author's account, the most significant release since the project launched. Reasoning models now stream their traces to standard error rather than standard output, so a piped invocation gets clean text while the thinking stays visible in the terminal, with a flag to suppress it. The release adds server-side provider tools through the OpenAI Responses API, redesigned content-addressable SQLite logging, and new models, with a matching llm-anthropic plugin update. The stderr-versus-stdout split is a small design decision that says a lot about building CLI tooling for reasoning models: the trace is for the human, the output is for the next process.

One model covering both the instruct task, where a caption specifies environment, speaker styles and fine-grained content with no reference recording, and the zero-shot task, where reference audio supplies the voice. The production motivation is concrete: dubbing, audio drama, advertising, games and short video all need voices designed from description and then reused consistently across a project, which is a persistence requirement most zero-shot systems do not address.

Five domains, each with 100 interconnected subtasks ordered by increasing difficulty and built to reward cross-task skill reuse, evaluating in-context continual skill learning rather than one-shot tool use. The headline finding is qualified: sequential execution generally improves performance, but the gains do not scale the way a compounding-skill story predicts. That gap between accumulation and compounding is the thing agent-memory products are currently selling and no benchmark had isolated.

A guest analysis of ChatGPT Work from Shlok, whose prior work reverse-engineering the memory systems of frontier labs is the reason to read it. The piece treats OpenAI's enterprise agent as a deployment problem rather than a model problem: what changes when an agent has to hold organizational context, respect permissions boundaries, and behave predictably across a workforce rather than a power user. Latent Space has been tracking OpenAI's agent deployments since Plugins in 2023, and the throughline of that coverage is that the hard part has consistently been retrieval and memory scoping rather than reasoning.

A geographic breakdown of where data center load growth is landing on residential electricity rates, which reached the Hacker News front page the same day Texas suspended new data center approvals pending audits. The causal attribution is contested in the comments and rightly so, since retail rates move on fuel prices, transmission investment and regulatory lag as much as on new large loads. What is not contested is the direction of the correlation, and the political salience of it, which is now shaping siting policy in at least two large states.

ARC has a returning executive director whose stated focus for the next six months is the organization's core research agenda: building techniques to find mechanistic explanations for neural network behavior, then using those explanations to detect and address misalignment. Jacob Hilton remains as VP of research. The post is candid that this is an ambitious bet attacking the core difficulty head-on rather than a portfolio play, and that government and developer advisory work is being scaled back to make room for it. For a field where most interpretability effort has drifted toward features and probes, an organization staking six months on explanations that are load-bearing for detection is a distinguishable position.

Optimization-based latent reasoning improves outputs by tuning instance-specific continuous states at test time with parameters frozen, but existing methods connect those states to the reasoning trajectory through decoded tokens, which makes sequence-level credit assignment indirect and hides how a latent update shapes what comes next. GradCuit places optimizable latent states at a selected transformer layer between the prompt's hidden representations and the generated continuation, so causal self-attention routes gradient directly through the circuit.

From the Inference Engineering Masterclass, a pointed argument that megakernels are dead: the case for spending two months hand-fusing a forward pass was launch overhead and poor inter-kernel overlap, programmatic dependent launch mostly closed that gap, and Rubin closes the residual straggler-CTA problem by letting a downstream kernel launch its ready CTAs while the upstream one drains. The blunt version, that no serious inference provider runs a 67,000-line hand-fused forward pass in production, is the sort of claim that gets tested in public quickly. The counter-position in the same discussion is that the megakernel idea survives as a compiler target rather than a hand-written artifact.

Existing query-only memory attacks stop working in the two settings that actually describe production systems: large benign memory pools, where a poisoned record has to win retrieval against real competition, and active input auditing, where it has to survive a semantic check. MAFIA combines probing to find retrieval-competitive phrasings with factual injection that reads as benign to an auditor. Memory-augmented agents are shipping faster than their threat models are being written, and this closes part of that gap.

An earlier paper on benchmark saturation resurfaced on the Hacker News front page, and its timing is apt given Vetto's Terminal Tasks release the same day explicitly citing Terminal-Bench 2.0 as saturating near 85 percent. The paper's contribution is treating saturation as a measurable regime with its own statistics rather than an anecdote about a leaderboard getting boring: once headroom compresses, score differences stop tracking capability differences and start tracking harness, prompt, and sampling noise. That is the failure mode every agent leaderboard is currently living inside.

Terminal user interfaces combine the stateful screen-oriented behavior of a GUI with terminal deployment and are everywhere in developer tooling, yet have no dedicated testing methodology. A survey of 197 real applications finds only 12 percent of test code exercises the interface at all, and 45 percent of those tests never send input, asserting on a static frame instead. The authors convert the applications into a headless benchmark across ratatui, bubbletea, textual and ink, packaged as instrumented Docker images with line and widget coverage recorded.

Jointly modeling languages with heterogeneous acoustic, phonological and lexical structure introduces optimization conflicts that erode per-language specialization, which is the standard tax on multilingual ASR. LS-MOPD decouples language-specific knowledge acquisition from multilingual integration by distilling from language-specialized teachers on policy, so the student sees specialist supervision on its own outputs rather than a compromise objective.

Models store a great deal of geographic knowledge and still fail at the geometric and topological computation that navigation and logistics require. MultiGlobeQA supplies 46,060 question-answer pairs across 14 spatial-function families and 15 answer formats with execution-based grading, multilingual and with controlled geographic coverage, replacing benchmarks that were synthetic, small, monolingual, or all three.

A BlueDot Technical Safety Project write-up that computes persona vectors and assistant axes for baseline models and for their emergent-misalignment model organisms, then uses geometric characterizations to separate two effects that the emergent-misalignment literature usually conflates: how badly a model behaves once it has adopted a role (behavior-per-role) versus how readily a given input elicits the misaligned role at all (role elicitation). Results are explicitly preliminary and the code is public. The decomposition is the useful part, because interventions that reduce elicitation and interventions that reduce per-role harm are different interventions and current evaluations score them together.

Merlin, which represents more than 30,000 independent labels and distributors, has joined Universal Music Group in backing Spotify's forthcoming AI remix and covers product. The paid tool will let listeners generate AI covers and remixes of participating artists' catalogs, with artists opting in, receiving credit, and being compensated. The licensing architecture is the substantive part: an opt-in, credited, compensated pathway is the first structure that plausibly scales generative music inside a major streaming service without litigating every training-data question first.

Bending Spoons has entered a definitive agreement to acquire Airtable for $1.285 billion, its first acquisition since going public. The Italian acquirer has built a business on buying mature software properties and running them for cash, and Airtable is a larger and more strategically awkward asset than its usual targets: a low-code database platform with a substantial enterprise install base and an unfinished repositioning around AI-assisted app building. The discussion on Hacker News focused less on the price than on what Bending Spoons' operating playbook typically means for a product's roadmap.

On-policy distillation presupposes teacher and student share VAE latents, architecture and timestep grid. When the best available teacher and the model you actually want to deploy come from different families, none of that holds: teacher latents are meaningless in the student's coordinate system and per-pixel losses against a stochastic teacher are noise. Any-OPD bridges in representation space instead, which is the only place the two models are comparable.

Ben Lorica's argument is that the standard production stack of benchmarks, thresholds and red teaming answers a narrow question well, namely whether a model is accurate, reliable, fast enough, and resistant to an adversary actively trying to break it, and answers almost nothing about what actually generates liability once the system is live. The cases he threads together, Workday, OpenAI, and a German court ruling, are all failures downstream of a model that passed its evaluations: process, disclosure, and jurisdiction problems rather than capability problems. The practical implication is that eval suites need a category for organizational and legal failure modes that no accuracy metric surfaces.

A long multi-turn agent trajectory may receive one outcome-level reward, so on-policy self-distillation is used to manufacture dense token-level supervision from a privileged teacher that is not reliable at every position. Existing weighting schemes either react to isolated token-level discrepancies, which is noise-sensitive, or apply one shared step-level weight, which ignores positional variation. PCSD weights on persistence of agreement across positions instead.

Handing an agent a skill library does not mean it can identify, apply and coordinate the skills in it. SKT builds skill-grounded tasks and executable trajectories from large skill collections, selecting single-skill and multi-skill configurations and synthesizing tasks through rule-based and agent-based verification. The verification step is what separates this from generic synthetic-data pipelines, and multi-skill coordination is the part current models are worst at.

Tool-use benchmarks expose semantic schemas in static environments, which lets an agent lean on prior knowledge instead of discovering how an unfamiliar system behaves. Removing the semantic cues and enforcing a continuous task curriculum isolates behavioral reasoning, and the title states the result: agents search exhaustively even when their own map points at the next step. That is a specific and testable failure of exploration policy rather than of reasoning capacity.

Multimodal on-policy distillation supervises student-generated trajectories with a privileged-view teacher, but the next-token corrections are source-mixed, blending genuine visual signal with linguistic priors and teacher-specific quirks. Visual Attribution Distillation is a counterfactual target-reconstruction algorithm that estimates the visually attributable component of each correction, so the student is trained on the part of the teacher that actually came from seeing.

Evaluating controllable video generators as world models has to go past appearance and explicit instruction following to inherent reactivity: whether the model can infer from scene state how the world should react and generate consequences nobody described in the prompt. Existing benchmarks check whether requested actions and stated outcomes were realized, which is precisely the part that does not require a world model. WorldExam is a hierarchical diagnostic aimed at the gap.

Most agent memory systems invoke additional LLM calls to write, summarize and retrieve records, adding recurring token and latency cost while merged or omitted details obscure the original evidence. Zero-Mem asks whether structured memory access requires generation at all, and answers no: nothing outside final question answering invokes an LLM or consumes LLM tokens, with encoder computation accounted separately. Whether it holds under adversarial retrieval is the open question, but the cost argument is straightforward.

In a new court filing Apple says its trade-secrets investigation into OpenAI has widened, claiming additional former staff may have retained or accessed confidential information. Talent-flow litigation between large technology companies is normally a cost of doing business, but the volume and seniority of movement between device makers and frontier labs has made the boundary between portable expertise and portable material genuinely contested, and this filing pushes on exactly that line.

Synthetic histopathology is pitched as an answer to data scarcity in computational pathology, but the evaluation apparatus is borrowed wholesale from natural images. The paper shows Frechet Inception Distance and Inception Score have real limitations on histopathology, where the perceptual features an ImageNet backbone extracts are largely orthogonal to the diagnostic features that matter, and proposes domain-specific metrics plus downstream task validation instead.

Most agent benchmarks use bounded tasks with immediate success criteria, which cannot measure the capacity to keep behavior purposeful across long horizons while adapting to accumulated evidence. Seller-side e-commerce is a good fit because actions constrain future choices, feedback arrives at heterogeneous delays, and incoherent behavior compounds measurably rather than failing loudly. The delayed-heterogeneous-feedback structure is the property that makes it hard.

Most AI systems emit an entire CAD program in a single pass and never inspect the intermediate geometry, while human engineers build a part feature by feature and check what remains after each operation. CADENA reconstructs a 3D mesh as a parametric CAD program by growing the operation sequence stepwise and comparing target against current predicted geometry at every step. Editability, not just shape match, is the goal that makes CAD reverse engineering worth doing.

Multimodal deep search has moved from single-turn factual retrieval to long-horizon multi-turn search guided by visual evidence, but existing methods confine vision to the input stage or the answer stage and skip its role in steering the search itself. DeepVoyager-VL incentivizes vision-in-the-loop, using visual evidence to decide what to look for next rather than only what to report at the end.

Feed-forward single-image 3D Gaussian splatting predicts Gaussians from fixed image-grid positions, which renders nearby views well but leaves the primitives weakly coupled to actual scene surfaces and falling apart under large viewpoint shift. InfiniSplat decodes Gaussians implicitly rather than pixel-aligned, targeting exactly the large-baseline regime where the pixel-aligned assumption breaks.

The latency-quality tradeoff in talking-head generation is structural: multi-step diffusion cannot stream, and real-time autoregressive methods accumulate error and drift in identity. LeapTalk's single-step bridge distillation aims to get one forward step per frame while scaling to arbitrarily long video, which is the combination that makes real-time avatar use cases viable rather than demo-able.

Learned sparse retrieval has stayed tied to encoder-style bidirectional architectures, and multimodal extensions lean on auxiliary cross-modal modules. UEmbed appends learnable special tokens to a decoder-only multimodal model and reads out both sparse lexical and dense representations from a single causal pass, which matters operationally because production search stacks want both and currently pay for two models.

A practitioner write-up on where AI-assisted programming has actually helped and where it has cost time, of the genre that reaches the Hacker News front page roughly weekly but is worth reading when it comes from someone working in a strongly typed, performance-sensitive codebase rather than a greenfield web project. The recurring finding across these accounts is that the productivity delta correlates less with model capability than with how much of the task is expressible as a well-specified local edit.

3D vision-language models build geometry-aware tokens by projecting 2D features into world coordinates, generating thousands of tokens per scene. Token compression developed for 2D relies on semantic relevance or attention-based selection and ignores the structured spatial character of 3D tokens, where redundancy is spatially organized rather than semantically distributed. 3DZip selects on spatial-aware feature diversity instead.

Clinical diagnosis, legal judgment and industrial fault diagnosis all require step-dependent causal chains where an early error propagates and a correct final answer can mask invalid reasoning. Trajectory imitation does not correct process errors on the student's own rollout distribution, so CausalOPD uses a curriculum of online process distillation targeted at the first wrong step. Locating supervision at the first divergence rather than spreading it across the trajectory is the efficient choice when errors are causally chained.

Endeavor Optical Networks plans to launch what it says will be the fastest space laser communications system yet built, aiming to move backbone traffic off subsea fiber. Free-space optical links have long had the theoretical bandwidth advantage and the practical problems of pointing, atmospheric attenuation, and weather diversity; the interest now is that inter-datacenter traffic for distributed training and inference is one of the few workloads with both the volume and the latency tolerance to justify an alternative path.

Database benchmarks remain fixated on text-to-SQL while the deployment story has moved to autonomous database administration. DBLifeBench spans five stages from schema design through post-deployment maintenance, which is where the operationally consequential failures live, and where a model that translates queries perfectly can still be useless.

Purely statistical causal discovery lacks the power to resolve structural ambiguity in low-sample regimes, and LLM-assisted hybrids improve recovery through semantic reasoning while leaving the influence of that reasoning on any given edge opaque. GENESIS targets the explainability requirement directly: why a specific edge is included in or excluded from the learned graph. For a method whose value proposition is scientific interpretability, that is the requirement that matters.

Most multilingual benchmarks test whether a model can perform a task in a language, conflating fluency with proficiency. M-GATE covers 30 typologically diverse languages from high to low resource across three tasks, starting with grammatical error detection on linguist-crafted adversarially selected sentences. Building the evaluation from linguistic judgment rather than translated task data is the methodological point.

AI infrastructure company Runware announced the Sonic Inference Pod, a modular data center intended to make inference capacity siteable wherever power and cooling are available rather than where a hyperscaler has already built. Given the week's other story out of Texas, where 474 gigawatts of interconnection requests have triggered a state audit, the argument for capacity that can follow stranded power rather than queue for grid connection is easier to make than it was a year ago. Whether pods clear the reliability and serviceability bar that inference workloads actually require is the open question.

Models routinely describe countries, organizations, historical events and social groups with affective framing layered onto the factual content: a target reads as favorable or threatening, calm or conflictual, powerful or vulnerable. Sentiment, favorability and emotion benchmarks each capture a slice; VIBE combines target-directed valence-arousal-dominance attribution with an explicit scorer contract and a standardized reporting format, which is what makes results comparable across models.

Video anomaly understanding requires identifying the abnormal event, finding supporting evidence and explaining the cause, which existing methods approach either with specialized training that does not generalize or with single agents that fold exploration, observation and decision into one reasoning pass. Separating explore from verify gives the system a place to reject its own evidence, which is where single-pass approaches quietly fail.

Captions supervise both multimodal understanding and text-to-image generation, and scoring them with a single scalar conflates how much visual information a caption covers with how reliably the image supports its claims. CAPEval decouples the two with human-written ground truth and human-verified atomic checklist items, which matters because the two properties trade off and the optimal point differs between the generation and understanding use cases.

Motion transfer normally assumes a fixed structural correspondence between reference and target, which is simply undefined when the two differ in morphology, articulation or deformation mechanism. The proposal is to transfer the dynamics that remain meaningful across morphologies via an abstract motion representation, in two stages. Cross-category transfer is the setting where the current formulation fails most visibly, so attacking the correspondence assumption directly is the right move.

Pluralistic alignment has become the standard response to the observation that a single unified value set cannot serve every deployment context, but current approaches lack an account of how values are socially organized, contested and coordinated in practice. The paper argues social theory supplies that account, and that without it pluralism collapses into a menu of preference profiles rather than a mechanism for coordination.

A Launch HN for EdotEnv, a Y Combinator S26 company building quantitative-trading reinforcement learning environments intended to train language models to do research rather than to trade. The interesting design claim is using markets as a verifier: a research hypothesis has a measurable out-of-sample consequence, which gives a dense and non-gameable reward signal of the kind that most agentic RL environments lack. The obvious failure mode is equally clear, which is that a model can learn to overfit a backtest as easily as a human can.

The Sci-ImageMiner dataset and ICDAR 2026 competition cover four end-to-end tasks in extracting knowledge from atomic layer deposition and etching figures, the kind of information that appears only in a plot and never in the paper text. Sixty-eight active participants produced 1,263 submissions. Expert-annotated scientific figure understanding is one of the few multimodal settings where the ground truth is unambiguous and the downstream value is obvious.

An analysis of seven years of Tesla earnings calls quantifies how much of Elon Musk's speaking time goes to robots and AI versus the car business. The result is a useful proxy for narrative-driven valuation in a company whose automotive fundamentals and whose autonomy-and-humanoid story have decoupled. Read alongside this week's US import ban on advanced foreign robots, it is a reminder that the domestic humanoid sector being protected is one whose most prominent participant has not yet shipped a product.

Existing Bengali text resources target handwritten documents or constrained signboard parsing, report only aggregate edit distance, and evaluate either conventional OCR or vision-language models but never both on the same in-the-wild data. BanglaWild supplies 2,535 scene text images with verbatim gold transcriptions, two categorical axes, four diagnostic attributes and an orthographically standard form where in-image text deviates from canonical spelling.

Industrial recommenders have adopted pretrain-then-transfer, which raises two coupled questions under behavioral drift: what to learn from behavior sequences, and how to transfer it while the pretrained model keeps being refreshed. Conventional next-token prediction treats adjacency as dependency and encodes spurious transitions across unrelated sessions; Behavioral Multi-Token Prediction retains only collaborative signal, and the geometry side handles refresh without retraining downstream.

Fixed-layout furniture styling must pick assets that form a coherent room without altering prescribed categories, positions, orientations or scales. Retrieving each asset independently or using static local relations produces shape, material and color conflicts once the scene is composed. A dynamic hypergraph style field with a frozen multimodal model supplying structured style priors makes the selection scene-level rather than per-object.