Wolf Digest — Tuesday, May 26, 2026

OpenAI announced this week that one of its internal reasoning systems produced a verified solution to an open Erdős problem in combinatorial geometry that had been unsolved since the mid-1940s. The Last Week in AI hosts gave the result roughly the same weight as DeepMind's earlier IMO and FrontierMath wins: a frontier model crossing a verifiable mathematical milestone that had resisted both human and prior-model effort. Independent verification of the proof is underway; the specific Erdős problem and the agent's full proof chain were not yet released at time of recording. The pattern — frontier labs racing to claim Erdős-era open problems as benchmark milestones — has accelerated noticeably over the past six months and is part of a broader shift in how labs publicly signal capability gains, away from chat-quality demos and toward verifiable mathematical wins.

Adversarial Flow Distillation (AFD) is an on-policy framework for distilling strong black-box video teachers into causal autoregressive students. The student rolls out under its own distribution while a prompt-paired Bradley-Terry discriminator scores teacher vs. student outputs on the same prompts; the resulting on-policy advantage is converted into forward-process flow-matching updates on the student's own noised states. The result: dense velocity-field supervision without requiring teacher scores, latents, denoising trajectories, step alignment, or reverse-chain RL. Two causal AR student families show consistent motion- and physics-sensitive improvements over off-policy SFT and prior adversarial baselines.

SkillOpt frames agent-skill improvement as a text-space optimization problem analogous to weight-space training: a separate optimizer model turns scored rollouts into bounded add/delete/replace edits on a single skill document, and edits are accepted only when they strictly improve a held-out validation score. A textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta updates make the procedure reproducible. Claimed as the first systematic controllable text-space optimizer for agent skills, with results showing consistent improvements on GPT-class agents under feedback signals where prior 'self-revision' baselines drifted.

Lens is a 3.8B-parameter text-to-image model that matches or surpasses state-of-the-art 6B+ models on standard T2I benchmarks while using only ~19.3% of the training compute Z-Image required. The efficiency gains come from two strategies: (1) Lens-800M, a dataset of 800M densely captioned image-text pairs whose captions average 109 words and are generated by GPT-4.1, providing richer per-batch semantic supervision than conventional short captions; and (2) constructing each batch from images with multiple resolutions to maximize information density per gradient step. The result is a compelling data-quality-over-parameter-count argument for the next generation of T2I models.

Claw-Anything expands the agent-evaluation surface along three dimensions current benchmarks neglect: long-horizon activity histories, interdependent backend services, and integrated GUI+CLI interaction across multiple devices. The benchmark simulates months of user activity via multi-round event injection to produce complex world states, then probes the agent's ability to act in that broader context. The framing matches the same week's product news (Gemini Spark, Perplexity Personal Computer) where 'always-on with broad access' is the explicit product target — so this benchmark has unusually high probability of becoming a standard reference for the category.

RTDMD is a two-stage framework unifying distribution-matching distillation with reward-guided RL for few-step flow generators. The authors show that minimizing KL divergence to a reward-tilted teacher distribution decomposes cleanly into a distribution-matching term and a reward-maximization term. Stage 1 (Ambient-Consistent Distribution Matching Distillation, AC-DMD) does subinterval-wise distribution matching with a consistency regularizer on the fake-score objective; stage 2 layers reward-guided RL on the distilled few-step generator. Empirically, the combined method aligns few-step diffusion models with human preferences more cleanly than either DMD or reward-only RL fine-tuning in isolation.

Channel-wise Vector Quantization (CVQ) replaces patch-wise tokens with channel-wise tokens: each channel of the feature map gets a discrete token, producing a representation of an image as 'discrete levels of visual detail' rather than a grid of spatial patches. A new visual autoregressive framework (Channel-wise Autoregressive, CAR) predicts image channels in sequence — sketching global structure first, then refining fine-grained attributes — and the resulting next-channel-prediction paradigm shows quality and efficiency gains over patch-AR baselines at matched parameter counts.

InstructSAM formulates instruction-driven instance segmentation as a set-structured query prediction problem with an explicit reasoning-to-instance query interface between a VLM and SAM3. A bank of learnable instance queries is injected into the VLM and contextualized with both instruction and visual tokens, and a hybrid attention mechanism promotes interaction among queries, vision tokens, and instructions to reduce duplicate predictions. The framework is competitive with task-specific segmentation pipelines while removing the need for closed instruction vocabularies.

AdvantageFlow optimizes a forward-process advantage-weighted prediction loss for rectified-flow models, in contrast to Flow-GRPO which optimizes the reverse process. The objective is unstable under negative advantages (non-convex), and the authors stabilize it via rollout-policy regularization derived from fitting a local reward-improving target distribution. Evaluated on Stable Diffusion 3.5 Medium image-generation tasks, AdvantageFlow outperforms both Flow-GRPO and a state-of-the-art forward-process negative-aware fine-tuning baseline.

DiscoverPhysics is an interactive benchmark with 22 simulated worlds whose laws of motion deviate from ours (screened/fractional-power gravity, multi-species couplings, hidden dark-matter-like particles, non-coordinate-free physics, time-varying interactions). An LLM agent proposes experiments, observes raw N-body trajectory data, and must submit both a natural-language theory of the world's physics and a Python implementation. The setup directly probes 'genuine reasoning vs. recall of established science' — a gap many physics-evaluation results have failed to disentangle. Frontier models, the authors report, do significantly worse than their established-physics evals would predict, suggesting most of their physics performance is recall.

LegalSearch-R1 pairs local statute RAG for precise article matching with online web search, trained end-to-end with RL on a temporal-consistency reward that punishes retroactive statute application — the bug the authors identify as the dominant failure mode of current legal LLM agents. Search agents in their evaluation rarely incorporate temporal constraints into queries, and base LLMs are anchored to training-cutoff law; the RL-trained pairing significantly improves both citation precision and time-appropriate-law selection on a held-out legal benchmark.

Foundation Protocol (FP) is a graph-first coordination layer unifying heterogeneous entities — agents, tools, resources, humans, institutions, organizations — and supporting native multi-party organization and event-based collaboration. It provides economic primitives (metering, settlement), identity/accountability primitives, and a discoverable namespace. The motivating thesis: as agents move from tools to social infrastructure, the bottleneck shifts from raw capability to coordination, and the field needs primitives at the protocol layer rather than ad-hoc per-platform glue.

DVAO targets a real failure mode in production RLHF/GRPO setups: standard scalarization (reward combination, advantage combination) either produces oversize squared advantages that destabilize training, or relies on static hyperparameters that ignore cross-objective correlations. DVAO dynamically adjusts the variance contribution of each reward component, producing more stable training and better multi-objective alignment than naive scalarization baselines on a suite of held-out evaluations.

VGenST-Bench actively synthesizes evaluation videos with generative models rather than relying on static curated clips, enabling controlled variation along a 3×2×2 taxonomy (spatial scale, perspective, temporal structure). A multi-agent generation pipeline with human QC stage maintains video quality; the benchmark probes spatio-temporal reasoning capabilities of MLLMs at a precision the curated-video benchmarks cannot match.

A study of how diffusion transformers route information across layers, identifying that most cross-layer signal flows through a small set of attention pathways that can be explicitly identified and pruned or strengthened. The authors propose a routing-aware training modification that produces measurable quality gains at matched compute by concentrating capacity along the discovered pathways.

QUEST is a family of open deep-research agents (2B–35B parameters) trained via a recipe combining mid-training, SFT, and RL on fully synthetic long-horizon search tasks. Strong claimed capabilities on fact-seeking, citation grounding, and report synthesis — the three subskills proprietary systems (OpenAI Deep Research, Perplexity Computer, Anthropic Skills) currently dominate. Open-weights release is the practically interesting bit; comparable to QUEST-class systems was previously only achievable behind APIs.

StepAudio 2.5 is a unified audio-language foundation model that matches or exceeds specialized systems across ASR, TTS, and realtime spoken interaction. The premise: once text and audio share a multimodal representational space, task specialization is a matter of operational regimes — different prompting, decoding, and conditioning — not architecturally distinct models. Concrete numerical wins are claimed against specialized baselines (WER on ASR, MOS on TTS, end-to-end latency on dialog).

Auto Benchmark Audit (ABA) is an agentic framework that systematically audits benchmark tasks, uncovering implicit assumptions, incomplete environment specifications, brittle evaluation logic, and incorrect ground truths. Run across 168 benchmarks (nine domains) drawn from frontier LLM evals and NeurIPS publications, ABA flags critical issues in over 25.7% of tasks — including ambiguous design, execution conflicts, and incorrect ground truths. The result is a concrete, large-scale rebuttal to taking benchmark scores at face value.

TriSplat is a feed-forward 3D reconstruction network that uses oriented triangle primitives instead of Gaussians, exporting simulation-ready meshes directly from a single forward pass on sparse views — no post-hoc mesh-extraction step. Works in pose-free settings (scene structure and camera parameters estimated jointly), making it suitable for downstream physics simulation and embodied interaction without the expensive conversion that breaks Gaussian-splatting's feed-forward promise.

WBench evaluates interactive world models along five dimensions: video quality, setting adherence, interaction adherence, consistency, and physics compliance. 289 test cases / 1,058 interaction turns spanning diverse scenes, styles, first-/third-person perspectives, and four interaction types (navigation, subject action, event editing, perspective switching). Lands the same week Google updated Genie to drive Street View — likely to be cited as the standard interactive-world-model eval in the next 6 months.

SciAtlas is a large-scale multi-disciplinary heterogeneous academic knowledge graph integrating 43M+ papers, designed to give research agents topological reasoning over scientific knowledge rather than relying on keyword or vector-similarity retrieval. The argument: vector-RAG-based research agents are prone to logical hallucinations and high inference cost; KG-grounded retrieval over a panoramic scientific evolution network produces qualitatively different (more grounded) deep-research output. Open release positions it as infrastructure for next-generation scientific-research agent stacks.

PolyGnosis 2.0 quantifies the efficacy of harness-engineering techniques (reflection loops, tool calling, divide-and-conquer partitioning, chain-of-thought) on extracting predictive signal from the gap between Polymarket sentiment and GDELT-derived OSINT — what the authors term 'Perspective Mismatches.' Empirically, divide-and-conquer harness structure dominates raw model intelligence on this task, reinforcing the system-scaling thesis (rank 21).

A position paper arguing that the next major bottleneck in agentic AI is system scaling, not model scaling: the design of auditable, persistent, modular, verifiable architectures around foundation models. Treats the harness — memory, retrieval, tool use, orchestration, verification, governance — as a first-class object of design and evaluation rather than 'secondary implementation detail.' Mirrors the empirical signal from Cursor Composer 2.5 above.

Macaron-A2UI argues static plain-text chat is the binding bottleneck for personal agents and presents a model family (30B/235B/754B, LoRA-fine-tuned) for generating natural language plus lightweight executable UI actions — information collection, preference refinement, confirmation, multi-goal organization — synthesized dynamically from interaction context. Trained on a large heterogeneous dialogue corpus with a dedicated A2UI-Bench evaluation; positions generative UI as a near-term necessary product layer for personal-agent products like Gemini Spark.

PiD reformulates the latent-to-pixel decoder of high-resolution T2I systems as a conditional pixel diffusion model — denoising directly in high-resolution pixel space — and unifies decoding with 4× and 8× upsampling in one generative module. Replaces the reconstruction-oriented decoder, which becomes prohibitively expensive at megapixel scale, with a more expressive synthesis-oriented one.

STORMS teaches LVLMs to reason through bounded continuous latent trajectories rather than externalizing reasoning via textual chain-of-thought, keyframe selection, or repeated frame reinsertion. Two-stage training internalizes the spatial-temporal reasoning, eliminating the inference-latency and engineering-complexity costs of the textual-CoT pipelines while matching or exceeding their accuracy on motion-tracking and temporal-order tasks.

ProAct is a proactive agent architecture that uses idle time between user interactions to predict and prefetch likely upcoming user needs, iteratively acquiring information and resolving knowledge gaps before the user initiates a query. Comes with ProActEval, a benchmark for proactive capability. The architecture maps onto the always-on assistant product surface and pairs naturally with Claw-Anything (rank 10) as the eval framework.

AnyScene generates semantic occupancy sequences from BEV layouts via a Spatial-Temporal Occupancy Diffusion Transformer that jointly tokenizes BEV and occupancy features autoregressively. Targets the long-tail safety-critical scenario problem in end-to-end autonomous driving simulation — replaces shallow conditioning and reference-frame-dependent video synthesis with arbitrary BEV-layout control. Pairs with Genie-on-Street-View (Google I/O) as the open-research version of the same world-model trend.

The Last Week in AI episode that consolidated the past week's news cycle: Google I/O 2026, the Cursor/xAI coding-agent race, Musk's lost OpenAI suit, Anthropic at $900B, Cerebras IPO, OpenAI's Erdős result, plus interpretability findings on redundant circuits, Terminal World benchmarks for agents, Take It Down Act enforcement on deepfakes, and demonstrations of autonomous hacking and self-replication. The episode is the most efficient single-source summary of the week's industry-side news; many items in this digest derive from threads it surfaced.

War on the Rocks essay arguing that AI-mediated automation is the keystone of any credible US economic-statecraft strategy to reshore manufacturing — and that current US industrial policy underweights the AI-deployment leg of the equation. The piece is positioned in the defense/national-security policy space and reflects a broader conversation about whether the AI capital surge will produce US-domiciled manufacturing capacity or whether it will replicate the offshoring dynamics of the 2000s in software instead of hardware.