Wolf Digest — Wednesday, June 3, 2026

Artificial Analysis published evaluation results for Alibaba's Qwen3.7 Plus, scoring it at an Intelligence Index of 53 with an AA-Omniscience of 2. That places it just below Kimi K2.6 at 54 and Gemini 3.5 Flash at 55, above NVIDIA's open-weight Nemotron 3 Ultra 550B at 48, and a clear step under the current leader Claude Opus 4.8 at 61. The result keeps Alibaba near the top of the closed-weight frontier and continues a dense run of Chinese-lab models, including DeepSeek V4 Pro, GLM-5.1, and Kimi K2.6, crowding the upper band of the Index.

K-BrowseComp is a 400-problem web-browsing agent benchmark grounded in Korean contexts, with a 300-problem human-verified subset. Frontier models including GPT-5.5, DeepSeek-V4-Pro, and GLM-5.1 reach only 30.0 to 45.67 percent, a steep drop from their BrowseComp scores, exposing how much agentic retrieval depends on language and cultural grounding rather than raw capability. Korean-tuned models close part but not all of the gap, and the benchmark stands as a reminder that agentic evaluations remain far from saturated outside English contexts.

Launched at Build, Microsoft Scout is an OpenClaw-inspired personal assistant that brings flexible, computer-use-style agency into the Microsoft 365 environment. It anchors a broader Build slate of agent tooling: portable policy files that let developer, compliance, and security teams define the rules an agent must follow, and an open-source spec-driven framework, Adaptive Spec-driven Scoring for Evaluation and Regression Testing, that spins up AI behavior tests from plain text descriptions. The throughline is governable, enterprise-deployable agents that can draw on internal company data, rather than another standalone chatbot.

Crafter is a multi-agent harness that generates editable, publication-quality scientific figures from diverse inputs rather than text alone, and unlike prior single-figure-type systems its outputs are locally revisable instead of flat raster images. By treating figures as structured compositions of discrete semantic components, Crafter can localize and repair the component-level errors generators make on such layouts. It topped Hugging Face Daily Papers for the window with 112 upvotes, the strongest community signal among the day's papers.

NVIDIA and Microsoft described a unified agentic-AI stack spanning Windows devices, Azure cloud, and local deployment, pairing NVIDIA's new Arm-based chips with secure runtimes and a data layer tuned for long-running reasoning. Microsoft's companion hardware, the Surface RTX Spark Dev Box, is a desktop built on those Arm chips for developers running models locally, while NVIDIA's NemoClaw push targets autonomous AI engineers for industrial computer-aided-design and simulation workflows.

This paper reframes parameter-efficient fine-tuning not as a cheaper substitute for full fine-tuning but as persistent per-instance local state layered on a strong shared base. Small adapters carry user preferences, skills, tool habits, and memory-like updates while the foundation model supplies shared competence. The authors organize the regime around scaling axes, Scale Up, where stronger shared priors make small updates more useful, and Scale Down, how small adapters can shrink while staying reliable, sketching a path toward millions of personal models layered over a trillion-parameter base.

On a Latent Space crossover recorded at Build, GitHub's Kyle Daigle laid out the platform's agent strategy: positioning GitHub as the control plane where humans and coding agents collaborate, with issues, pull requests, and Actions becoming the substrate for autonomous work rather than human-only workflows. The conversation surfaced the central tension of GitHub as a neutral collaboration surface versus the agent harnesses, Claude Code, Codex, and Devin, each racing to own the developer's primary loop.

Domino is a speculative-decoding framework that decouples causal-dependency modeling from expensive autoregressive draft execution. Rather than choosing between slow autoregressive drafters, which model intra-block dependencies but incur sequential overhead, and fast parallel drafters, which are cheap but weaken dependency modeling, Domino drafts in parallel and uses a separate lightweight pass to capture token dependencies, recovering draft quality without the sequential cost. It targets the central latency bottleneck in LLM serving.

Nathan Lambert announced his departure from the Allen Institute for AI, where he led work on the open Olmo models. His reflection argues that Olmo's impact came less from frontier performance, which it never reached even within size buckets, than from fully open training recipes, data, and tooling that set norms for reproducible open science. The post doubles as a meditation on diverging paths to impact in AI today: chasing frontier capability versus building open infrastructure and shared community knowledge.

A new bill in Congress aims to set guardrails on military uses of AI, arriving the same week as Trump's deregulatory civilian AI order and creating an unusual split: a lighter touch for commercial frontier models, tighter scrutiny for defense applications. Reporting situates it within an active legislative season for AI oversight, alongside open questions about autonomy, meaningful human control, and procurement as the Pentagon accelerates its drone and autonomy programs.

TASTE, Task Synthesis from Tool Sequence Evolution, reverses the usual agent-benchmark construction pipeline. Instead of writing natural-language scenarios and mapping them to tool calls, which captures only a narrow slice of real tool-use patterns, TASTE evolves tool sequences first and synthesizes harder tasks from them, automatically broadening coverage and difficulty. It targets the saturation of benchmarks like tau-squared-Bench as agent capabilities climb past them.

Harness-1 is a 20B-parameter retrieval subagent trained with reinforcement learning inside a state-externalizing harness. The insight is that forcing the policy to manage routine bookkeeping, what has been seen, which evidence matters, which constraints remain open, which claims were checked, wastes RL capacity on recoverable state the environment can track more reliably. By externalizing that state, RL optimizes the genuinely semantic search decisions, improving long-horizon retrieval over policies that carry everything in context.

H Company released Holo3.1, a family of fast computer-use agents designed to run locally for GUI control. The pitch is latency and privacy: small enough to run on local hardware while handling screen grounding and multi-step interface actions, positioning open computer-use agents as an alternative to cloud-only systems for desktop automation.

This study maps when masking stale observations from a search agent's context actually helps. Sweeping backbones from 4B to 284B parameters and three retrievers across offline and live-web benchmarks, the authors find the accuracy gain follows an asymmetric inverted-U: helpful up to a point, then harmful as useful evidence gets dropped. The paper offers a mechanism for the regime boundaries, giving practitioners a principled rule for context-budget management rather than a blanket heuristic.

TrOPD targets the regime where on-policy distillation breaks: when teacher and student distributions differ substantially, teacher supervision on student-generated tokens yields unreliable gradients and can cause optimization to fail. It applies trust-region credit assignment to token-level supervision, making OPD usable for agent learning, multi-task enhancement, and model compression where the student strays far from the teacher.

Next Implicit Token Prediction augments standard next-token prediction with dense continuous supervision directly in the representation space. The argument is that sparse one-hot label supervision under-constrains hidden states, letting them drift into degenerate, anisotropic configurations that limit generalization; NITP additionally trains the model to predict the implicit semantic content of the next token, regularizing the latent space during pretraining.

This paper shows distributional LLM watermarks are fragile under linear ensembling: averaging logits across a few models, or mixing a watermarked model with unwatermarked ones, substantially removes the statistical signal that watermark detectors rely on, with little quality cost. The result complicates provenance and attribution schemes that assume watermarks survive routine model mixing, a relevant caveat as policymakers lean on watermarking for AI-content disclosure.

Inspired by human memory consolidation, this Sleep paradigm lets a model continually learn by distilling short-term, in-context knowledge into long-term parameters during dedicated offline phases. It targets the gap where LLMs handle instant prediction and in-context learning well but cannot transfer temporal, in-context knowledge into their weights, a step toward continual learning without catastrophic interference.

Humanoid-GPT scales data and architecture for zero-shot whole-body motion tracking on humanoids, training a generalist controller that follows reference motions it never saw during training. The emphasis on data scale and structural priors continues the foundation-model trend in legged control, where broad pretraining is starting to yield zero-shot generalization to novel motions.

Ben Thompson argues Alphabet increasingly behaves like a capital-allocation company, deploying its balance sheet through large equity raises and external investments to fund the compute and infrastructure buildout that defines the current AI cycle. The framing reframes Google's competitive position around capital deployment and infrastructure ownership rather than product alone, echoing the broader shift toward compute as the binding constraint.

The Information frames a crowded enterprise-AI sales moment: both OpenAI and Microsoft pressed new business pitches this week, role-specific agents, forward-deployed engineers, and homegrown models, at corporate buyers already overwhelmed by competing AI vendors. The piece raises the practical question of differentiation and integration cost when every major lab is simultaneously courting the same technology-buying decision-makers.

X-Stream studies using multimodal LLMs as multiplexers that interleave and route several simultaneous input streams, for example multiple video or sensor feeds, through one shared model. The work analyzes how attention allocates across streams and where capacity bottlenecks appear, probing whether a single model can hold multiple live contexts at once for multi-stream understanding.

This work uses vision-language models as teachers to improve video reasoning through adaptive test-time optimization, distilling VLM judgments into a video model at inference and adapting per-instance. It improves temporal reasoning without retraining the base video system, trading a modest inference-time cost for better reasoning on long or complex clips.

VideoMLA adapts multi-head latent attention to video, introducing a low-rank latent key-value cache that compresses the memory footprint of minute-scale autoregressive video generation. The result enables longer rollouts at lower VRAM while preserving temporal coherence, attacking the memory wall that limits how long autoregressive video models can run.

This paper develops a local perturbation theory for why RL post-training on one domain, such as math, code, question answering, or creative writing, often degrades the others even when full-model gradients do not globally conflict. The framework localizes where interference and recovery happen, offering a more precise account than catastrophic forgetting or global gradient conflict for the cross-domain trade-offs seen in multi-domain RL fine-tuning.

LiveBand generates high-fidelity musical accompaniment in real time to live audio input under strict causal constraints. A causal transformer generator operates in the continuous latent space of a pretrained causal audio autoencoder, trained with sequence-level adversarial supervision, and at each timestep receives only the causally available mix context, a step toward AI that can back a musician live rather than in offline post-production.

An empirical study comparing pretraining paradigms for downstream spatial-intelligence tasks finds that the objective best serving spatial reasoning differs from what best serves semantic tasks. The implication is that spatially-grounded foundation models may need pretraining objectives tuned for geometry and viewpoint rather than borrowed wholesale from language-style or contrastive recipes.

AutoMedBench is a workflow-aware benchmark for agentic medical-AI research that evaluates not just final outputs but the intermediate steps agents take across end-to-end research workflows. By scoring the process, it exposes where autonomous research agents go wrong in retrieval, hypothesis formation, and analysis, rather than only grading the conclusion, addressing a blind spot in prior medical-agent evaluations.

Uber capped employee AI spending after exhausting its budget within a single quarter, a small but telling data point on how fast internal AI-tool consumption is scaling and why enterprises are now imposing cost controls on agentic tools that bill per token or per task. It echoes a recurring theme this week of companies trying to keep AI bills in check as usage outruns forecasts.

Google rolled out on-device fake-call detection to flag AI-generated voice deepfakes during phone calls, part of a wave of consumer-facing defenses against voice-cloning scams. The feature lands as synthetic-speech quality increasingly outpaces listeners' ability to detect cloned voices unaided, pushing detection down to the handset.

Greece expanded its fleet of Shield AI V-BAT autonomous drones for maritime operations, extending the company's Hivemind autonomy stack into a NATO member's naval intelligence, surveillance, and reconnaissance mission. The deal underscores continued international procurement of AI-piloted uncrewed systems for contested maritime environments.

Data-security startup Cyera is eyeing a $12 billion valuation at roughly 80 times annual recurring revenue despite operating losses, a marker of how richly investors are still pricing AI-adjacent security companies. The multiple lands amid broader scrutiny of whether AI-era revenue can grow into valuations set well ahead of profitability.

Martin Scorsese, a longtime skeptic of AI in filmmaking, became the latest and most unlikely Hollywood figure to engage with generative tools, a cultural signal of shifting industry attitudes as AI video generation matures. The move reads less as endorsement than as a sign that even the medium's traditionalists now feel compelled to reckon with the technology.