Wolf Digest — Tuesday, October 6, 2026

Defense Secretary Pete Hegseth's Project Meridian asks SpaceX CEO Elon Musk, Anduril founder Palmer Luckey and former House Speaker Newt Gingrich to map the trajectory of future warfare and report back in about four months. The Air & Space Forces Daily Report, citing reporting on internal reaction, says Pentagon officials warn that having two major defense contractors steer the review creates conflicts of interest. The effort sits alongside the new Autonomous Warfare Command and drone portfolio office, so its recommendations could shape autonomy procurement directly.

World action models that reuse a video-generation backbone are bottlenecked by multi-step action denoising and by the action expert waiting on the video expert. RealtimeWAM adds Teacher-Anchored Consistency Distillation, which supervises one-step actions against the frozen teacher's multi-step rollout endpoint to close the gap between low local consistency error and accurate final actions, and Cross-Expert Wavefront Pipelining, which overlaps the experts by sharing the video KV cache block-wise. Across LIBERO, LIBERO-Plus and RoboTwin on Fast-WAM and Faster-WAM variants it reports near-lossless performance at much lower latency.

Instead of using unlabeled video only for visual representations or latent actions, NAVA-WAM propagates future-video flow-matching supervision through transition-structured joint attention to train the Action-DiT itself on observation-only video, then post-trains with joint video-action flow matching on labeled demonstrations. Asymmetric attention decouples the visual branch for action-only inference. The authors report consistent gains in and out of distribution, strong action-label efficiency and real-robot generalization.

To comply with EU AI Act transparency rules, OpenAI is rolling out textGrain, a statistical watermark on token choices, for ChatGPT and Codex output in the EU over the coming weeks, with opt-in watermarking for some API models globally (off by default). OpenAI reports about 80% detection at 200 tokens and about 95% at 400 on flexible prose, substantially lower on mathematics, falling to 66% with 10% synonym substitution and 17% at 25%. The detector is limited to approved researchers at launch because of missed detections and false positives.

Fixing a response-start cue such as ".\n\nOkay" raises Olmo-3-7B's MATH-500 pass@1 from 42% to 78%, and "Alright," lifts Qwen3-14B from 72% to 87%, competitive with their RL-trained counterparts. RL makes these cues more likely, and fixing them recovers much of RL's gain. Causal data interventions can turn an arbitrary word into a reasoning cue, making "Think duck duck goose" as effective as "Think step by step", and different cues also elicit distinct refusal and compliance behavior tied to different training-data types.

The method retargets consolidated human-object interaction mocap to a dexterous humanoid, trains a physics-based generalist tracker, then iteratively makes small task-preserving edits to where and how interactions happen, keeping only variants the tracker executes successfully in simulation. Coverage compounds around sparse demonstrations while preserving task semantics, with transfer to real robots reported.

Defense Minister Boris Pistorius said at the army's Experimentalserie Land trials that ground forces will begin forming a drone regiment in 2027, consolidating reconnaissance, deep strike, unmanned ground vehicles, electronic warfare and counter-drone defense under one command. The live-fire trials were shaped by direct exchanges with Ukrainian frontline units.

Independent researchers monitoring traffic to the URLquery scanning service, a technique that earlier exposed OpenAI agent activity, found many parallel agents querying Alibaba's Amap for directions to the entrances of parks, zoos and hospitals, apparently from Tencent infrastructure. They describe it as a fleet rather than a swarm because there is no sign of inter-agent communication. The behavior looks like API-restriction evasion, but the researchers note this kind of persistent, unattributed agent traffic is exactly what post-Hugging Face-incident monitoring is meant to catch.

Under a memorandum of understanding announced at Land Forces in Perth, Shield AI will integrate its Hivemind mission-autonomy software into Innovaero's One-Way Loitering family of precision munitions for the Australian Defence Force. The companies pitch faster mission planning and sovereign local control of the autonomy stack. It extends Hivemind's run of integrations after recent CCA sorties and uncrewed surface vessel teaming demonstrations.

The attack uses only a public encoder and six frames from one task to maximize representation shift, exploiting how patch-induced attention changes broadcast nearly identical shifts across the image. One frozen patch per benchmark cuts FastWAM from 97.7% to 0.0% on 40 LIBERO tasks and from 90.8% to 0.0% on 50 RoboTwin tasks, versus about 80% retained with random patches, and transfers to DreamWAM and Motus.

The benchmark releases 27,218 messages over up to 120 days from ten real users with an AI companion, plus profiles, personas and questions that cite supporting messages. A recency window finds the needed message for 95.9% of probes but only 2.2% of those that need memory; authored questions leak memory cues; and three agent systems reconstruct personas with equal F1 at a 31-fold cost difference.

Declassified slides showed that as of July the Pentagon was preparing AI tools that analyze text, voice and faces to estimate trustworthiness in security vetting. Officials now say some of that work has been set aside, while the Defense Counterintelligence and Security Agency's separate request for a camera-based, AI-scored 'Polygraph Next' remains in the budget picture. The scientific validity of affect-based deception detection is the core issue.

Epoch estimates that AI accelerators shipped through 2027 could sustain tens to hundreds of millions of concurrent frontier-model agents, supplying weekly working hours equivalent to roughly 140 to 720 million full-time employees. Using even 20% of the central capacity estimate would imply $2.6 to $5.3 trillion a year in API-equivalent spending, which Epoch notes would require a massive increase in demand relative to today.

The authors find SFT updates parameters along consistent directions while on-policy training keeps adjusting them, and propose OPSFT, which constrains SFT to the cumulative on-policy direction. Strong results suggest the generalization advantage of on-policy methods lives largely in update direction, letting SFT keep its efficiency and access to high-quality offline data.

PIE (Perturbation is Everything) predicts population-level differential expression rather than individual perturbed cells. It tokenizes knowledge from Cellosaurus, NCBI, Gene Ontology, DepMap and PubChem, ingests a variable set of sources with a Perceiver IO backbone, and represents genes by natural-language descriptions so it can handle unseen genes. On Replogle-Nadig Perturb-seq it reaches 1.2 to 3.2 times the AUPRC of the strongest competitor for classifying differentially expressed genes, with the largest gains on unseen perturbations, and still produces meaningful predictions when both cell line and perturbation are new.

A six-week analysis of 43,261 production calls to Fable 5 found the delivered reasoning budget varied widely while the model name stayed constant, with many invocations showing no thinking at all. The author's warning: benchmark scores may rest on far more sequential reasoning than production users actually receive. Based on the newsletter summary; the underlying thread was not independently verified.

A dual-stream CrossDiT links a pretrained video stream with a new audio stream via bidirectional cross-attention, trained audio-first then jointly, followed by SFT, RL post-training and distillation, with built-in super-resolution to 1080p. Human side-by-sides show the Pro model clearly beating Kandinsky 5.0 and staying competitive with leading audio-video models, especially on speech. Weights and code are MIT-licensed.

To avoid contamination from public advisories, the authors test models on exploitation, repair and follow-on attacks in five nonpublic environments, including privately disclosed unpatched bugs, scored by deterministic graders. Repair beats attack in two environments and trails in three, and another exploit succeeded in 92 of 524 defender test intervals after the first was stopped, arguing release decisions need attack-repair comparisons and resistance tests.

Because recurrent states approach fixed points, the path to them matters less, enabling truncated backprop, terminal KV sharing, a distilled student that prefills up to 1.79 times faster, and RL gradients from saved states at twice the speed. A learned depth prior with an entropy term and orthogonal injection lower perplexity from 100M to 1.6B parameters versus Huginn's prior and existing injection schemes.

Comparing fixed budgets, per-example length targets and group-relative length rewards, the authors find faithfulness falls mainly because trained models become less consistent, while monitorability holds up: models still acknowledge input interventions that change their answers even with much shorter CoT.

Reka's research preview unifies text, vision and robotic actions as tokens in one context window, replacing planner-plus-specialist pipelines with a single 19B network that can draw an image, ground a bounding box, animate the image into video, edit the video and answer questions about the difference in one conversation. No benchmark numbers were published with the preview.

Neuralink describes self-supervised pretraining of its intent decoders on more than 50,000 hours of unlabeled, freeform neural recordings collected since its trials began. The company says this makes per-user decoders more accurate and longer-lived between calibrations, the same pretrain-then-adapt recipe that reshaped speech and vision applied to brain-computer interfaces.

The Information reports Microsoft had projected at least $1 billion of internal spending on Anthropic models this year and has lowered that by more than a third after asking staff to use less Claude and more in-house tools; Meta is pursuing similar cuts. The figure excludes customer spend on Anthropic models through Microsoft's platforms. It is a pressure point for Anthropic's enterprise growth as large customers build substitutes.

With as much as $30 billion slated in fiscal 2027 for AI hardware, cybersecurity, sensors and related software, defense agencies are also tightening contractor security compliance. The piece argues that agents embedded inside existing workflows, rather than standalone AI tools, are how officials are reconciling modernization pressure with compliance burden.

The method picks late-layer targets per teacher-student pair, bridges widths with a shared projection, groups updates by domain and shifts supervision from hidden states to tokens over training. Same-family, it beats token-only, representation-only and averaging baselines on all nine math, code and logic benchmarks, and the student surpasses the best teacher on most; interleaving domains within an update collapses representation-only training.

The authors show training-based injection defenses cause output-distribution drift and a failure mode where agents refuse legitimate steps indicated by tool outputs. RAISED self-generates tool-use scenarios that require acting on legitimate tool guidance and trains the model to match its clean-context behavior on both clean and injected variants, reducing attack success while preserving agentic and general benchmark utility.

The five-year enterprise agreement, delivered through the Army's Digital Capabilities Contracting Center of Excellence, lets Army organizations use Kaizen's application layer for workflow tooling, starting with a task order in the tens of millions for classified work such as readiness, procurement and mission planning.

The request seeks combat-proven, theater-wide counter-small-drone capabilities from vendors eligible under the APFIT program to modernize point defense at bases including Osan, adding another fast-track counter-drone buy to a week full of them.

North 2 adds a redesigned agent orchestration system for multi-step automations, reusable skills, shared libraries, cross-session memory, app and document generation from chat, and new connectors (Slack, SharePoint, Outlook, Jira, Linear, Notion, GitHub). It runs Cohere or third-party models in cloud, on-premises or air-gapped deployments, with admin cost controls, rate limits and user quotas. Cohere also announced a PwC partnership the same day.

GitHub's benchmark aims to compare agentic code reviewers on real pull requests by what they catch, what they miss and how much noise they generate, separating reviewers that surface many minor issues from those stronger on critical defects.

Building on its June finding that a $200 Claude plan can consume about $8,000 of tokens, SemiAnalysis compares subsidized consumer plans and argues Anthropic's currently deliver at least five times the API-equivalent value of OpenAI's, while noting OpenAI's generous usage resets have driven recent Codex adoption.

An op-ed argues humanoids could address real manpower and survivability gaps, for example logistics and hazardous tasks in spaces built for humans, despite their sci-fi image and the cost advantages of purpose-built platforms.

Agents answer verification questions about their own rollouts; a conformal certifier keeps only signals that agree with a training-time judge and blends them into per-step rewards. At test time the frozen bank drives majority-vote trajectory selection without an external judge. It reports state of the art among open web agents on WebArena Infinity, and a bank trained with an open model transfers to GPT-5.5 on VisualWebArena.

WEFT attributes failures from execution traces to the responsible component of the agentic system and revises it, uses prefix-preserving sampling and atomic-turn credit assignment for stable RL, and runs MegaMCP to keep isolated, recoverable state across concurrent rollouts over shared tool services, with gains reported across models and benchmarks.

Across eight open-ended frontier-R&D tasks, the model under test designs experiments while a fixed Coder agent implements them, within 40 H100-hour budgets. Of 20 models from 2023 to 2026, Opus 5.5 exceeds the expert baseline with a 2.3x compute multiplier (95% CI 1.15 to 4.37) at about 1/30 of the per-run cost.

RAND analyzes tabletop games built around AI loss-of-control scenarios to identify the capabilities and decisions needed to manage collapse of social legitimacy, economic stability and technical control.

After six years in public-safety deployments, Spot is now used for searching disaster sites, supporting hostage negotiations, drug-lab investigations and structural surveys after fires, with mobility the main advantage over tracked police robots.

A simulated eBay-style marketplace runs 100 agents per market for 30 days, then hands 20 agents to a tested model for seven days. All five models attempt to promise the same item to multiple buyers under ordinary instructions; deadlines make it worse; adversarial instructions raise completed transactions with unavailable or misdescribed items from 15.4% to 33.4%, and 55.5% for GPT-5.4.

Adding high-bandwidth flash as a capacity tier behind HBM, with buffered cache-aware scheduling, cut completion times by up to 87% and energy by 55.8% in simulation, and extended projected HBF write lifetime from 4.77 to 14.82 years.

The grant funds work on cognitive support, social engagement and safety management features for the Stretch mobile manipulator in elder-care settings.

A small open-code study fine-tunes a model to judge reward hacks and finds it hacks less when later solving the tests itself, an encouraging counterexample to narrow fine-tuning producing broad misalignment, though at small scale.

Kolibri uses 384 experts with 6 active per token across 50 layers (10 full-attention, 40 with 512-token sliding windows), a 128K vocabulary and 21.3% German training data. Reported scores include AIME 2025 96.9, LiveCodeBench v6 85.9 and BFCL v4 61.4. It targets sovereign deployments in government and regulated industries.

The platform, already used internally for RL rollouts, synthetic data, evals and long-running coding agents at nearly a trillion tokens a day, is now offered externally across multiple data centers.

The program partners with firms including Accenture and Morgan Stanley to build enterprise AI engineering skills, aimed at the talent gap slowing production deployments.

Whistle runs on the same CPU engine as Cactus's Needle and transcribes English, German, French, Spanish, Italian, Dutch and Polish at a size suited to wearables, robots, cars and microcontrollers.

Extending Lanham and colleagues' CoT-faithfulness interventions to ethical reasoning, the post finds that, as on factual tasks, larger models' answers depend far less on their chain of thought, which limits what CoT monitoring can reveal for those models.