Wolf Digest — Tuesday, May 12, 2026

DECO is a sparse Mixture-of-Experts architecture targeting end-side deployment that activates only twenty percent of experts and matches dense Transformer performance at identical total parameter budgets and training tokens. The routing uses a differentiable ReLU gate with learnable expert-wise scaling and a new activation function called NormSiLU that normalizes inputs before SiLU to produce a more stable routed-expert activation ratio. The authors also identify an empirical advantage in using non-gated MLP experts with ReLU routing, suggesting MoE architecture simplification is possible. A specialized acceleration kernel delivers a three-times speedup on real hardware versus dense inference. Code and checkpoints will be released.

Flow-OPD ports on-policy distillation from LLM post-training to Flow Matching text-to-image models, addressing the reward-sparsity and gradient-interference issues that produce the 'seesaw effect' in multi-task FM alignment. The two-stage recipe trains domain-specialized teachers via single-reward GRPO, then consolidates them into a single student via Flow-based cold-start plus on-policy sampling, task-routing labeling, and dense trajectory-level supervision. Manifold Anchor Regularization uses a task-agnostic teacher to anchor generations to a high-quality manifold and prevent the aesthetic degradation typical of RL-driven alignment. On Stable Diffusion 3.5 Medium, Flow-OPD raises GenEval from sixty-three to ninety-two and OCR accuracy from fifty-nine to ninety-four.

SLIM treats the active external skill set of an LLM agent as a dynamic optimization variable jointly updated with policy learning, rejecting the assumption that skills either accumulate as persistent guidance or get fully internalized into the policy. The method estimates each skill's marginal contribution via leave-one-skill-out validation, then applies three lifecycle operations: retain high-value skills, retire skills whose contribution has become negligible after sufficient exposure, and expand the skill bank when persistent failures reveal missing capability coverage. Experiments show an average 7.1-point improvement over the best baselines on ALFWorld and SearchQA, with some skills absorbed into the policy and others continuing to provide external value.

Microsoft Research released SocialReasoning-Bench, an evaluation framework for measuring whether AI agents act in users' best interests in multi-party negotiation, allocation, and bargaining settings. The headline finding from the launch post: across the models evaluated, agents execute the requested actions competently but consistently fail to improve the user's outcome relative to baseline, even when explicitly instructed to optimize for user interest. The benchmark fills a gap between pure capability evaluations (which measure whether the agent can do the task) and pure preference evaluations (which measure whether users like the result), targeting the principal-agent alignment failure that arises in practical deployments.

Anthropic's Claude Opus 4.7 release, which has been the model behind the new Artificial Analysis Coding Agent Index top scores this week, is now the lead Anthropic frontier model on the Intelligence Index at fifty-seven — tied with Gemini 3.1 Pro Preview and GPT-5.4 xhigh and one point behind GPT-5.5 xhigh. Anthropic frames the release around stronger performance on coding, agents, vision, and multi-step tasks, with the model showing particular thoroughness and consistency on long-horizon work. Listed on Anthropic News this week as the top product card, the model is the one driving Cursor CLI's and Claude Code's wins on the new Coding Agent Index.

DeepSeek's V4 Preview release, captured in this run's browser sweep, is now live and open-sourced with one-million-token context as the default across all official services. DeepSeek-V4-Pro has 1.6 trillion total parameters and 49 billion active; DeepSeek-V4-Flash is 284 billion total / 13 billion active. The structural novelty is token-wise compression combined with DeepSeek Sparse Attention (DSA), which the team credits with the dramatic context-length efficiency. On Artificial Analysis's Intelligence Index, DeepSeek V4 Pro Max sits at fifty-two, alongside Claude Sonnet 4.6 and Muse Spark. V4 is integrated with Claude Code, OpenClaw, and OpenCode out of the box. The legacy deepseek-chat and deepseek-reasoner endpoints will retire on July 24.

A pointed result for deep-research-system builders: a well-configured BM25 lexical retriever paired with frontier LLMs reaches 83.1% answer accuracy and 94.7% surfaced-evidence recall on BrowseComp-Plus, outperforming released search agents that use dense retrievers. The Pi-Serini agent uses three tools — retrieve, browse, read — and shows that BM25 tuning alone yields a 18.0-point answer-accuracy improvement over default BM25, and increased retrieval depth gives a further 25.3-point evidence-recall improvement over shallow retrieval. The implication for current deep-research agent stacks is that the dense-vs-lexical retrieval split has been overweighted; capable agentic loops can extract substantially more from lexical retrieval than the SOTA-chasing dense literature implied.

PhyGround evaluates physical reasoning in video generation with 250 curated prompts, each paired with an expected physical outcome, organized under a 13-law taxonomy across solid-body mechanics, fluid dynamics, and optics. Each law is operationalized through observable sub-questions for per-law diagnostics. The authors ran a quality-controlled human study with 459 annotators producing 5,796 complete annotations and over 37,400 fine-grained labels across eight video generators; split-half model-ranking correlation exceeded 0.90. The release includes PhyJudge-9B, an open physics-aware automated evaluator. The benchmark's structural contribution is operationalizing physical laws as observable sub-questions rather than holistic plausibility ratings.

G-Zero bypasses the proxy-LLM-judge bottleneck for self-evolving LLMs in open-ended tasks. The core innovation, Hint-delta, is an intrinsic reward that quantifies the predictive shift between the Generator's unassisted response and its response conditioned on a self-generated hint. A Proposer trained via GRPO continuously targets the Generator's blind spots by synthesizing challenging queries and informative hints; the Generator is concurrently optimized via DPO to internalize these. The authors prove a best-iterate suboptimality guarantee for an idealized standard-DPO version, conditional on the Proposer inducing sufficient exploration coverage and on pseudo-label score-noise being small.

Step Rejection Fine-Tuning (SRFT) leverages SWE-bench trajectories that standard Rejection Fine-Tuning would discard. A critic LLM assesses correctness step by step, and during training the loss is masked for erroneous steps while keeping them in the context window — so the model learns to recover from errors without reproducing them. On SWE-bench Verified, vanilla RFT lifts resolution rate by 2.4 points by excluding unresolved trajectories; SRFT lifts it by 3.7 points by retaining them with step-level masking, reaching 32.2% total resolution. The finding is a concrete recipe for the standard trajectory-curation problem in coding-agent post-training.

Project Glasswing is the joint AI-security initiative across AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks aimed at securing the world's most critical software. Surfaced in the browser sweep of Anthropic News this run, the alliance's mandate framing centers on AI-mediated security review of widely used open-source and infrastructure code. The structural contribution is the multi-company coalition rather than any single Anthropic product — the listed members span hyperscalers, financial infrastructure, and standards bodies.

Reinforce Adjoint Matching is an RL post-training method for diffusion and flow-matching generators that targets the gradient-stability issues that have made standard policy-gradient methods unreliable on diffusion backbones. The cross-source coverage (five arXiv categorical feeds picked it up) reflects a methodology paper sitting at the intersection of RL, generative media, and evaluation tooling rather than a single-axis contribution. The framing is on scaling the RL post-training of these models without the variance blow-up that has limited prior approaches.

Mira Murati's Thinking Machines is publicly signaling work on a model that processes input concurrently with generating output — eliminating the strict turn-taking that every contemporary LLM relies on. The TechCrunch piece frames the architectural problem: today's LLMs treat input and output as serial, and conversational latency plus the inability to revise output mid-stream are both downstream effects of that constraint. The company has not released technical details; the article is a positioning piece for what Thinking Machines is hiring toward. The framing matters because the streaming-input/streaming-output regime is the natural home for voice agents and live interactive use cases.

Jack Clark's Import AI 456 covers three threads: a theoretical look at the interaction between recursive self-improvement of AI systems and economic growth modeling, an argument for 'radical optionality' as a regulatory posture in fast-moving AI environments, and coverage of a recent neural-computer paper. Clark's editorial framing on radical optionality argues that policy frameworks designed to lock in specific safety outcomes underperform frameworks designed to maximize the regulator's future option-space — relevant context for the ongoing policy debates around capability evaluations and pre-deployment review.

Stratechery's subscriber-only Monday piece argues that the AI value-capture story is shifting from training-side to inference-side economics — and that chip companies positioned for inference workloads have an unusually favorable IPO window in May 2026. The full text is behind the paywall, so this summary is necessarily light, but the framing pairs directly with this week's OpenAI-Microsoft restructuring and the $10 billion private-equity vehicle: both moves read as OpenAI repositioning around an inference-dominated future economic structure rather than the training-dominated one that defined the 2023-2025 deal architecture.

A joint Hugging Face / Amazon-AWS post enumerating the building blocks AWS exposes for foundation-model training and inference: SageMaker HyperPod for distributed training, Inferentia and Trainium silicon, and the Bedrock-side integration paths for serving HF-hosted models. The post is positioned as a reference for ML teams choosing between vertical providers and DIY stacks on AWS primitives. Practical takeaway: AWS's pitch is to give teams the cluster-management and chip-level flexibility of Trainium while keeping the high-level deployment workflows familiar from Bedrock.

OpenAI's Q1 2026 usage update reports that ChatGPT adoption surged in the first quarter with the fastest growth among users over 35, and a more balanced gender split than in prior quarters. The framing — broader mainstream adoption rather than the previously concentrated under-35 tech-worker base — is consistent with OpenAI's Personal Computer and enterprise narrative this quarter. The post is short and avoids absolute MAU numbers, which limits independent verification, but is the cleanest first-party signal available on demographic shifts.

OpenAI's enterprise playbook codifies the migration from early experiments to compounding AI impact through trust, governance, workflow design, and quality at scale. Aimed at chief AI officers, the guide tracks the gap between pilot and production deployments and recommends specific governance and quality-at-scale practices. Daniel-adjacent context: the playbook is downstream of the same enterprise-pull narrative that the OpenAI-Microsoft revenue-share rework relies on for its top-line growth model.

General Motors laid off hundreds of IT staff and is hiring replacements focused on AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, prompt engineering, and new AI workflows. The framing is one of the cleanest current data points on the displacement-versus-augmentation question inside large traditional employers — the company is explicitly replacing roles with different roles requiring different skill sets, not eliminating headcount overall.

AssayBench operationalizes the virtual-cell vision as a phenotypic-screen prediction task: 1,920 publicly available CRISPR screens spanning five broad classes of cellular phenotypes, framed as gene-rank prediction per screen with an adjusted nDCG metric that allows comparison across heterogeneous assays. The empirical headline is that zero-shot generalist LLMs outperform biology-specific LLMs and trainable baselines on this benchmark — a striking inversion of the usual expectation that domain-specialized models win on domain-specific tasks. Existing methods remain far from the empirically estimated performance ceilings; fine-tuning, ensembling, and prompt optimization further close the gap.

The Defense Innovation Unit's latest news, surfaced this run via Chrome MCP capture: Space-BACN, the satellite optical-communications program for inter-satellite laser links, formally transitions from DARPA to DIU on May 4, 2026. The shift signals that the program is moving from research-prototype phase to acquisition. Other DIU items in the same window: five additions to senior leadership (April 17), Buckley SFB and Malmstrom AFB selected for the Advanced Nuclear Power for Installations program (April 8), and a Defense One piece on aircraft losses and the demand for software-driven battle-space awareness.

The Missile Defense Agency is targeting a 2027 demonstration for Project Maverick, a hypersonic-weapon interceptor. The acquisition timeline is one of three new DefenseScoop dispatches this morning, alongside the new S&T chief's update on the DoD pilot program for free tech patent licenses and a profile of the Strategic Capabilities Office's incoming director.

Mistral AI, surfaced in this run's Chrome MCP capture of mistral.ai/news, is named as a founding member of the NVIDIA Nemotron Coalition, contributing large-scale model development and multimodal capabilities. The coalition is NVIDIA's industrial alliance around the Nemotron model family; Mistral's inclusion is the strongest single signal that the European open-weights player is leaning further into NVIDIA's commercial silicon-and-model stack.

Embedded Language Flows formulate compact text generation as a continuous flow-matching problem on token embeddings, sitting alongside the growing wave of diffusion-style decoders for language (LLaDA, Mercury) but using continuous flows rather than discrete masking. The cross-category sourcing (cs.AI + cs.CL + cs.LG + generative media) reflects a methodology that bridges the LLM and image-generation literatures.

Cohere Transcribe, surfaced in this run's Chrome MCP capture of the Cohere blog, is a new open-source speech-recognition model positioned as a state-of-the-art result in the ASR space. Cohere's framing emphasizes the open-source release rather than enterprise-only API access, which would be a departure from Cohere's typical commercial posture if borne out in the release artifacts. The release lands in the same week as ElevenLabs's $500M ARR announcement, suggesting continued commercial momentum in the speech-and-audio segment.

ElevenLabs disclosed crossing $500 million in annual recurring revenue and welcomed new investors including BlackRock and NVIDIA, alongside celebrity participants Jamie Foxx and Eva Longoria. The disclosure is notable for the ARR scale — among the highest of any TTS or speech-focused company — and for the BlackRock entry, which is rare among AI-tooling cap tables. For the Wolf Digest specifically, ElevenLabs is the TTS provider behind the daily audio narration of this digest.

Stability AI's Brand Studio, surfaced in this run's Chrome MCP capture, is an end-to-end creative-production platform pitched at enterprise marketing functions. The product wraps Stability's image, video, audio, and 3D models behind brand-conditioned controls — visual identity, type, and copy guidelines — to enforce consistency across automated generation. Strategically it follows the Warner Music and Universal Music partnerships from late 2025 and the EA partnership; Stability is repositioning as an enterprise platform play rather than an open-model lab.

Apollo Research, the alignment-and-evaluations lab behind a number of public scheming-and-deception evaluations of frontier models, converted to a Public Benefit Corporation in January 2026. The PBC structure is the conversion path used previously by Anthropic and OpenAI's restructuring proposals; Apollo's framing is that the PBC structure aligns better with its mission of alignment evaluations than the standard for-profit corporate form. The blog post surfaced in this run's Chrome MCP capture of apolloresearch.ai/blog.

The May 2026 Transformer Circuits drop, surfaced this run via the Chrome MCP capture, introduces Natural Language Autoencoders — a method that trains Claude to translate its internal state into natural-language explanations. The headline is unsupervised explanation generation for LLM activations, sitting in the lineage of the Sparse Autoencoder feature dictionaries but producing language-level outputs rather than learned feature codes. The release ships alongside HeadVis, an interactive visualization tool for attention-head behaviors. Both are interpretability-team output and reflect Anthropic's ongoing investment in mechanistic-interpretability tooling.