Wolf Digest — Saturday, May 23, 2026

Artificial Analysis posted a busy changelog this week. New language-model evaluations: Grok 4.3 medium (22 May), Grok 4.3 low (21 May), Gemini 3.5 Flash minimal (21 May), Gemini 3.5 Flash high (19 May), Qwen3.7 Max (19 May), Ring-2.6-1T (19 May), Command A+ (20 May). New image-model entries: Recraft V4.1 and Recraft V4.1 Utility (21 May). On Intelligence Index v4.0 the leader board reads GPT-5.5 xhigh 60.2, Claude Opus 4.7 max 57.3, Gemini 3.1 Pro Preview 57.2, GPT-5.4 xhigh 56.8, Qwen3.7 Max 56.6, Gemini 3.5 Flash 55.3, Kimi K2.6 53.9, MiMo-V2.5-Pro 53.8, Grok 4.3 high 53.2, with DeepSeek V4 Pro Max at 51.5 and GLM-5.1 at 51.4. The Coding Agent Index puts Claude Code w/ Opus 4.7 (max) at 67, Codex w/ GPT-5.5 (xhigh) at 65, and Cursor CLI Composer 2.5 Fast at 63 — the last is the price story Latent Space flags separately.

MIT Tech Review’s I/O readout argues the more interesting signal than Demis Hassabis’s “foothills of the singularity” line was that most substantive examples in the keynote were drawn from science rather than consumer products. Google rolled out Gemini for Science, a researcher-facing suite that bundles long-context analysis of papers and datasets, multimodal interpretation of figures and microscopy, and a co-scientist agent that proposes hypotheses, designs experiments, and critiques drafts. The shift Tech Review names is from one breakthrough model per domain (the AlphaFold playbook) to general-purpose research assistants — AI Co-Scientist, Isomorphic Labs’ drug-design models, Weather Lab’s nowcasting, AlphaProof — that compress literature review, code, and experimental design into a single conversational interface. Stanford geneticist Gary Peltz, quoted from a Nature Medicine piece, compared Co-Scientist to “consulting the oracle of Delphi.” Critics in the article note persistent hallucinated citations and out-of-distribution struggles, plus a real risk of scientific monoculture if every lab consults the same model.

NVIDIA’s Nemotron Labs releases a family of diffusion-based language models targeting much faster text generation than the standard autoregressive setup. Instead of producing one token at a time conditioned on all prior tokens, the diffusion path generates many tokens in parallel by iteratively denoising a sequence over a small number of refinement steps, trading the strict left-to-right factorization for parallelism that maps cleanly onto modern accelerators. The flagship model was pre-trained on 1.3T tokens from NVIDIA Nemotron Pretraining and SFT-tuned on 45B Nemotron Post-training tokens; architecturally it keeps a transformer backbone but swaps the causal mask for bidirectional attention during denoising and uses a discrete diffusion objective adapted to tokens. Reported wins: competitive quality on standard reasoning/coding benchmarks plus substantial latency and throughput improvements that widen at long output lengths. Deployment is supported through SGLang with optimized kernels; weights drop on Hugging Face under a permissive license. The framing positions diffusion LLMs as a credible second track alongside autoregressive scaling, most attractive for latency-sensitive agent loops and code-completion.

Dharma AI argues that once training data is moved close to the deployment task, parameter count stops being decisive. Evidence: a 3B specialized OCR model topped a curated DharmaOCR-Benchmark of real enterprise documents (invoices, contracts, scanned ledgers, mixed languages), beating Claude Opus 4.6, GPT-5.5, Gemini 3 Pro and several other frontier APIs at roughly fifty times lower cost per page. The gap to Claude Opus 4.6 — close to eight points — is wider than any other adjacent gap in the table, which the authors take as evidence that specialization produces a step-change rather than incremental gains. Pipeline: continued pre-training on domain text, supervised fine-tuning on task-specific input-output pairs, RLAIF with a stronger judge, and synthetic-defect augmentation (noise, skew, low contrast, mixed languages). Strategic argument: in domains with stable distributions and clear success criteria — document processing, claims adjudication, vertical customer-support triage, single-codebase code review — a properly fine-tuned small model can dominate on cost/latency while matching or beating frontier APIs on quality.

Linear-attention models compress an unbounded softmax cache into a fixed-size recurrent state, but the open question is how to edit that state without scrambling existing associations. Delta-rule models subtract the current read before writing; Kimi Delta Attention (KDA) sharpens forgetting with channel-wise decay. Both still use a single scalar gate to control two different things: how much old content to erase on the key side and how much new content to commit on the value side. Gated DeltaNet-2 separates these roles with a channel-wise erase gate b_t and a channel-wise write gate w_t, reducing to KDA when both gates collapse to the same scalar and to Gated DeltaNet when the decay also collapses. The paper derives a fast-weight update view and a chunkwise WY algorithm with channel-wise decay absorbed into asymmetric erase factors. At 1.3B parameters on 100B FineWeb-Edu tokens, Gated DeltaNet-2 is the strongest among Mamba-2, Gated DeltaNet, KDA, and Mamba-3 across LM, commonsense, and retrieval, with the largest advantage on long-context RULER needle-in-a-haystack.

RTPurbo argues full-attention LLMs are already intrinsically sparse and can be retrofit into highly sparse models with only minimal adaptation. Three observations: only a small subset of heads truly needs full long-context, long-range retrieval is governed by a low-dimensional subspace (a 16-dim indexer suffices), and the useful token budget is strongly query-dependent (dynamic top-p beats fixed top-k). The method retains the full KV cache only for retrieval heads and adds a lightweight token indexer for sparse attention, hitting sparsification with only a few hundred training steps. On long-context benchmarks and reasoning tasks, RTPurbo preserves near-lossless accuracy while delivering up to 9.36× prefill speedup at 1M context and ~2.01× decode speedup — a strong argument that you don’t need native sparse pretraining to get the inference wins.

Standard agent SFT masks tool responses and only trains turn-level tool selection — a supervision blind spot for the scattered evidence inside long trajectories. ACC (Agent Context Compilation) converts trajectories from search, software-engineering, and DB-query agents into long-context QA pairs that combine the original question with tool responses and environment observations gathered across many turns, training the model to answer directly without tool use. Training Qwen3-30B-A3B with ACC hits 68.3 on MRCR (+18.1) and 77.5 on GraphWalks (+7.6), matching Qwen3-235B-A22B while preserving GPQA/MMLU-Pro/AIME/IFEval. Mechanism analysis finds task-adaptive attention restructuring and expert specialization. A clean recipe for scaling long-context supervision without new human annotation.

CUSP (Cutoff-conditioned Unseen Scientific Progress) evaluates 4,760 scientific events across feasibility, mechanistic reasoning, generative solution design, and temporal prediction under controlled knowledge cutoffs. Frontier models can identify plausible directions from candidates but fail to reliably predict whether advances will be realized and systematically misestimate timing. Performance is highly domain-dependent (AI timing more predictable than biology/chemistry/physics) and is largely insensitive to whether events fall pre- or post-cutoff, suggesting limitations cannot be explained by training-data exposure alone. Models exhibit systematic overconfidence and strong response biases. Bottom line: today’s systems are unreliable as forward-looking forecasters of scientific progress, and access to prior knowledge does not translate into reliable forecasting.

A Cogs of War interview with Catarina Buchatskiy (Snake Island Institute) and Viktoriia Honcharuk on how Ukraine’s defense ecosystem converts battlefield feedback into rapid iteration. End-user requirements drive innovation: tech that reaches the front is evaluated immediately under harsh conditions, in contrast to Western development where requirements often originate far from the user. The interviewees describe platoon commanders sending 15-second failure videos to manufacturers over Signal and receiving software patches within hours. Companies that succeed locate manufacturing in or near Ukraine, push OTA updates, and embed engineers with units. Frontline R&D labs prototype, test, and iterate at the edge, then hand mature designs to larger manufacturers for scale — an example cited is Dopkhin’s Pavuk long-range drone (50+ km behind lines). They push back on the “Ukraine = drones” framing: the ecosystem now spans counter-UAS, EW, ground robotics, logistics autonomy, and battlefield AI. The implicit advice for Western partners: stop selling finished platforms and start learning the iteration cadence.

Disaggregated LLM serving (PD separation, KV state disaggregation) turns KV cache into an explicit payload crossing network and storage boundaries, making it a dominant end-to-end bottleneck. Existing KV compression is static; KVServe unifies KV compression into a modular strategy space, adds a Bayesian Profiling Engine that distills a 3D Pareto candidate set (cutting 50× offline search overhead), and deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under SLO and bandwidth constraints. Integrated into vLLM and evaluated across datasets, models, GPUs, and networks, KVServe achieves up to 9.13× JCT speedup in PD-separated serving and up to 32.8× TTFT reduction in KV-disaggregated serving — the kind of number that matters when KV egress is the real bill.

Process Reward Models give step-level supervision but typically require expert annotations per step. uPRM trains PRMs without human supervision — no step-level annotations and no ground-truth final-answer verification — via a scoring function derived from next-token probabilities that jointly assesses candidate positions of first-error steps across a batch of reasoning trajectories. uPRM hits up to +15% absolute accuracy over LLM-as-Judge on ProcessBench, matches supervised PRMs as a test-time-scaling verifier (and beats majority voting by up to 6.9%), and produces more robust RL policy optimization than a supervised PRM trained with ground-truth labels. A meaningful step toward scalable reward modeling for complex reasoning.

TerminalWorld reverse-engineers high-fidelity evaluation tasks from 80,870 in-the-wild terminal recordings, yielding 1,530 validated tasks across 18 categories (short ops through 50+ step workflows, 1,280 unique commands), plus a Verified subset of 200. Across eight frontier models and six agents, the best pass-rate is 62.5%. Pearson correlation with the curated Terminal-Bench is only 0.20, suggesting real-world terminal capability is a distinct axis from what expert-curated benchmarks measure — an important caveat for anyone using Terminal-Bench scores as the headline number for shell agents.

LatentOmni argues that for joint audio-visual reasoning, explicit text-based CoT compresses continuous sensory signals into discrete tokens, weakening temporal grounding. It interleaves textual reasoning with audio-visual latent states, adds feature-level supervision to align those latent states with task-relevant sensory features, and uses an Omni-Sync Position Embedding (OSPE) for temporal consistency. The authors release LatentOmni-Instruct-35K, a dataset of interleaved reasoning trajectories. Across multiple audio-visual reasoning benchmarks LatentOmni is the best open-source result evaluated and consistently beats the explicit text-CoT baseline, evidence that latent-space joint reasoning is a serious path for omni-modal understanding.

π-Bench targets proactive assistance — the ability to anticipate user needs that are never explicitly stated. The benchmark spans 100 multi-turn tasks across five domain-specific personas, with hidden intents, inter-task dependencies, and cross-session continuity that exercise long-horizon proactivity rather than single-turn task completion. Headline findings: proactivity remains hard, task completion and proactivity are clearly separable axes (good on one often bad on the other), and prior-interaction memory meaningfully helps with proactive intent resolution downstream. A useful complement to the current crop of single-task agent benchmarks.

DIU and the Army announced a joint Driverless Cars Prize Challenge on 22 May, the latest in a string of DIU prize-style procurements (Autonomous Vehicle Orchestrator, Blue Object Management, Counter-UAS Low-Cost Sensing) that bypass traditional acquisition. The mechanism is the news as much as the topic: DIU is using prize authority to pull commercial autonomy talent into ground-vehicle work for the Army at speeds the program-of-record pipeline can’t match. Pairs with DIU’s 4 May announcement that the Space-BACN satellite-laser-link program transitioned from DARPA to DIU and DIU’s 19 May project spotlight on XR training devices — the autonomy and edge-AI tempo at DIU is up.

Cohere’s 20 May release positions Command A+ as a sovereign-agentic open-weights model — the first major Command-series update since Command A more than a year ago. Artificial Analysis added a Command A+ evaluation the same day. Same week Cohere also announced strategic MOUs with Indra Group (defense electronics) and Multiverse Computing, and the Reliant AI acquisition to expand into biopharma/healthcare — the company is leaning hard into the sovereign-enterprise positioning that competitors with closed weights can’t directly match. Falls just outside the strict 24-hour window but cross-references back into today’s industry threads.

A 13M-record dataset of public-transit route plans across four Chinese cities (120,845 stations, 13,666 lines) released as a continual-pretraining corpus and a benchmark for three evaluation tasks. An LLM trained on TransitLM produces structurally valid routes and implicitly grounds arbitrary GPS coordinates to appropriate stations without any explicit mapping infrastructure, suggesting end-to-end map-free transit planning is feasible from data alone. Useful pretraining substrate for spatial-reasoning work even outside the transit setting.

Stability releases Stable Audio 3.0, a model family pitched explicitly for artistic experimentation and built on fully licensed training data — a deliberate response to the IP-litigation pressure that has dogged Stability and Suno over the last year. Open weights are part of the framing. Comes on the heels of the WMG and UMG partnerships from late 2025/early 2026 that set up the licensing rails to make a clean-data audio model commercially defensible. Just outside the strict window but worth recording because it’s the headline audio-gen release of the week.

The Information reports Microsoft is pushing on a new agent-data front — the contested question of which web data, enterprise data, and partner data AI agents are allowed to traverse, cache, and act on. The dispute matters because the agent-economics story (Latent Space’s thesis above) only works if there’s a stable substrate of data the agents can actually use. Pairs with the GitHub Gartner-Leader story and CATL’s reported DeepSeek investment for a clean snapshot of where the agent-industry stack is consolidating.

CATL, the world’s largest EV battery maker, is reportedly preparing to invest in DeepSeek. The capital matters less than the signal: a Chinese industrial-strategic anchor backing the country’s most prominent open-weights frontier lab consolidates DeepSeek’s role inside China’s broader sovereign-AI buildout. Lands the same week DeepSeek’s API docs flipped over to V4-flash/V4-pro as primary models and Artificial Analysis recorded DeepSeek V4 Pro Max on the Intelligence Index at 51.5 (top of the open-weights cluster).

GitHub announces Gartner recognized it as a Leader in the inaugural Magic Quadrant for Enterprise AI Coding Agents, the same quadrant that now formally includes Copilot, Cursor, Codeium/Windsurf, Cognition’s Devin, and others. Substance is less important than the institutional signal — enterprise procurement now treats “AI coding agent” as a category requiring its own analyst quadrant, with implications for budget cycles at every Fortune 500 software org. Reads alongside the Latent Space coding-agent thread (Composer 2.5 at 3rd, frontier-lab vs. specialist economics).

DefenseScoop reports Space Force is accelerating on-orbit logistics work — the docking, refueling, and servicing infrastructure that determines whether US satellites are persistent assets or single-launch-per-mission consumables. Robotic autonomy and AI scheduling are core enabling components. Same news cycle as DIU’s announcement that the Space-BACN satellite-laser-link program shifts from DARPA to DIU.

PhysX-Omni unifies generation of rigid, deformable, and articulated 3D objects in a single model with simulation-ready outputs. Aimed at robotics and simulation pipelines that need geometry plus physical parameters (mass, friction, joint limits) coming out of the same diffusion call. Notable as part of the trend toward "physical 3D generation" rather than texture/mesh-only output — inputs that drop into MuJoCo or Isaac Sim without manual cleanup.

Spreadsheet-RL applies RL to spreadsheet agents on realistic tasks (multi-sheet workbooks, formula chains, charts), targeting the gap between toy spreadsheet benchmarks and what enterprise users actually do. Sits inside the broader agent-RL wave — the same training-time-RL story the labs above are building products around.

WorldKV proposes an efficient world-memory module for embodied / long-horizon agents that combines retrieval and compression to keep relevant context queryable over long episodes. Approach overlaps the recurrent-memory and KV-compression literature but is targeted at agentic settings where the "world" is multi-modal and persistent across turns.

An extension of the Platonic Representations Hypothesis to human neural data, claiming unsupervised methods recover the same universal geometry across human brain recordings as has been found across deep nets. If it replicates, it lends weight to the strongest version of representation universality — that biological and artificial systems converge on the same conceptual geometry given enough capacity and data.

Dwarkesh’s second conversation with Reiner Pope on what it actually takes to design AI chips from the bottom up — process node selection, memory hierarchy decisions, the interaction between compiler stack and silicon, and the asymmetric advantages NVIDIA holds in the inference market. Companion piece to the broader narrative about Cerebras, Groq, and the new wave of inference-specialized silicon.

The Marine Corps tested using a rotorcraft as a mobile command-and-control node for tactical drone operations — essentially treating the helicopter as a flying command post that brings the human operator close enough to the swarm to overcome bandwidth and latency limits of distant control. Concept of operations matters: small-unit air-ground integration with autonomy in the loop is a USMC modernization priority that aligns with the DAWG framing.

SEGA addresses the resolution-extrapolation problem in diffusion transformers — generating at resolutions higher than training — by guiding attention with a spectral-energy criterion. Promises higher-resolution outputs without quality degradation typical of naive extrapolation.

DeepMind announces SynthID, its imperceptible watermark for AI-generated content (text, image, audio, video) is expanding to more partners — the practical question for the watermark-vs-deepfake arms race is whether enough of the generation surface area is covered to make detection useful at scale. Each new partner narrows the unwatermarked frontier.

Methods for fine-tuning and post-training interactive diffusion music models for live performance: low-latency adaptation, controllable generation, and stable behavior under continuous user interaction. Lands in the same week as Stable Audio 3.0 — the live/interactive corner of audio generation is heating up.

Sensor2Sensor proposes a learned conversion between sensor modalities/configurations across AV platforms, letting models trained on one stack transfer to another without re-collecting data. The cross-embodiment transfer story — a big theme in robotics this year — ported to the AV setting.

SOCOM is fielding a new tactical biometrics system for identification at the point of contact. AI on the device side handles match-on-device against curated watchlists. Civil-liberties critics will note this is the same vendor ecosystem that has been pitching state and local law enforcement on derivative products; the operational deployment by US special operations will likely accelerate procurement debates in domestic contexts.

TechCrunch’s hands-on of Google’s AI glasses: form factor and on-device speech are good enough to feel inevitable, but battery, display quality, and ambient privacy still aren’t. Useful corrective to the I/O keynote framing.

TechCrunch dissects how AI-startup ARR figures are constructed today: annualized monthly run-rates, bundled platform credits, partner co-marketing commitments, and one-off enterprise pilots all rolled into the headline. Reads alongside the ElevenLabs $500M ARR / new investor announcement and the OpenAI revenue stories swirling around the IPO — a healthy skepticism filter on the current revenue-acceleration narratives.