Wolf Digest — Tuesday, June 16, 2026

Z Lab, Modal, and SGLang released DFlash, a speculative-decoding draft model that generates an entire block of draft tokens in parallel via block diffusion and injects target-model latents directly into every draft layer's KV cache, rather than autoregressively drafting like EAGLE or native MTP. The released DFlash drafter for Qwen 3.5 397B-A17B beats both baseline and native MTP in every tested setting, reaching over 4.3x baseline throughput at concurrency 1 on HumanEval (8xB200). Paired with SGLang's new default Spec V2 overlap scheduler, end-to-end throughput rose ~33% (11.4 to 15.3 ktok/s on Qwen 3-8B, single B200, concurrency 32). Drafters released across three Hugging Face orgs.

Qualcomm has been in talks to acquire Tenstorrent, the AI-chip startup founded by Jim Keller, at a price discussed between roughly 8 and 10 billion dollars, a steep premium to Tenstorrent's last valuation, per The Information. It is unclear whether the figure includes milestone-tied payments common in chip-startup acquisitions, and the talks could still change price or fall apart. A deal would expand Qualcomm's AI accelerator capabilities and add Tenstorrent's RISC-V-based architecture and training/inference designs to a portfolio historically centered on mobile and edge silicon.

Data2Story orchestrates specialized agent roles into a single virtual newsroom to take raw data to a finished, trustworthy news feature end to end. Two design choices stand out: an Inspector grounds every claim, number, angle, and asset back to data, code, or an external reference, and articles are multimodally generative, reasoning about which visuals and layout best convey each point rather than defaulting to plain text and static charts. The paper drew 70 HF upvotes, reflecting strong interest in verifiable, end-to-end agentic generation of long-form analytical content.

Contrary to expectations that cloud providers' custom server chips would erode Nvidia's dominance in inference, The Information estimates Nvidia's share of the inference-chip market rose to about 74% from 66% over the past year, based on company disclosures and analyst interviews. Inference, running already-trained models, had been seen as the segment most exposed to in-house accelerators from hyperscalers, making the gain a notable signal that Nvidia's software stack and supply continue to hold share even where alternatives are most viable.

VibeThinker-3B probes how far verifiable reasoning can be pushed in a strictly small-model regime. Built on the Spectrum-to-Signal post-training paradigm with curriculum SFT, multi-domain RL, and offline self-distillation, the 3B dense model reports 94.3 on AIME26 (97.1 with claim-level test-time scaling), 80.2 Pass@1 on LiveCodeBench v6, and a 96.1% acceptance rate on unseen LeetCode contests, placing it in a performance band usually associated with far larger reasoning models. The result adds to evidence that careful post-training, not just scale, drives verifiable-reasoning gains.

TokenPilot tackles the tension between trimming an agent's accumulated context and preserving prompt-cache continuity: text pruning and dynamic eviction shrink token footprints but mutate sequence layouts, causing prefix mismatches and cache invalidation. Its dual-granularity framework pairs Ingestion-Aware Compaction, which stabilizes prompt prefixes and filters environmental noise at the ingestion gate, with Lifecycle-Aware Eviction, which offloads context segments only when their task relevance expires on a conservative batch-turn schedule. Surfaced across six sources, the highest cross-source count of the day's papers.

Qwen-RobotWorld is a language-conditioned video world model that predicts physically grounded future visual trajectories from current observations across robotic manipulation, autonomous driving, indoor navigation, and human-to-robot transfer, using natural language as a unified action interface. It couples a frozen Qwen2.5-VL's semantics with video-VAE latents through a 60-layer double-stream MMDiT with layer-wise joint attention, trained on an 8.6M video-text Embodied World Knowledge corpus. The unified formulation targets three uses: synthetic data for policy training, scalable virtual environments for evaluation, and language-guided planning signals for control. Six sources.

FastContext separates repository exploration from problem solving in LLM coding agents, where locating relevant code normally burns token budget and pollutes the solver's context with irrelevant snippets. Invoked on demand, it issues parallel tool calls and returns concise file paths and line ranges as focused context. The exploration models span 4B-30B parameters, bootstrapped from strong reference trajectories and refined with task-grounded rewards for broad first-turn search and multi-turn evidence gathering. Drew 35 HF upvotes, reflecting active interest in context-efficient software-engineering agents.

Salesforce has agreed to buy Fin, the customer-service agent startup formerly known as Intercom, for 3.6 billion dollars, a premium to Fin's last estimated 1.8 billion dollar valuation, per The Information. The deal is part of Salesforce's push to win enterprises onto its own AI offerings and signals continued consolidation in the customer-support-agent market as incumbents acquire purpose-built agentic products rather than build them in-house.

Sen. Ruben Gallego sent Defense Secretary Pete Hegseth a letter, obtained by DefenseScoop, pressing the Pentagon to disclose how it is addressing operational risks to U.S. personnel as officials work to meet the administration's new 90-day mandate to rewrite policy for deploying and safeguarding autonomous weapon systems. Gallego asked specifically about measures to mitigate possible unintended harm to Americans and allies that could accompany a rapid revision of the long-standing directive governing autonomy in weapons.

Tencent has invested 20 million dollars in a new AI lab founded by Junyang Lin, former lead researcher on Alibaba's Qwen models, according to The Information's sources. The investment is part of a first round that raised several hundred million dollars; a related report indicates the lab is seeking a valuation around 2 billion dollars. The move signals continued churn of senior frontier-model talent in China into well-capitalized new ventures, with major platform companies backing breakaway labs.

CODA-BENCH is presented as the first benchmark to jointly evaluate code and data intelligence in a data-intensive environment, closing a gap left by benchmarks that test code-centric or data-centric ability in isolation. Built on a Kaggle-derived Linux sandbox with hundreds of datasets, it requires agents to actively explore complex file hierarchies, identify relevant resources, and generate code for data-driven analytical tasks across 1,009 tasks. The design more closely mirrors real development work, where agents must navigate both large codebases and large file systems.

StepPO argues that token-centric RL inherited from RLHF and RLVR mismatches the granularity of agentic decision-making, where LLM agents act in steps, cycles of environmental observation and action, not tokens. It reformulates agentic RL from a token-level MDP into a step-level MDP, treating interaction steps as the basic trajectory unit, and introduces step-level credit assignment to align optimization with that natural granularity. The work joins a cluster of recent papers reworking credit assignment for multi-turn agentic training.

MIT Technology Review profiles Casey Harrell, who has ALS and has lived with implanted electrodes for nearly three years, first using his brain-computer interface to 'speak' sentences with a research team in 2023 and since logging thousands of hours. He now operates the device largely independently once a carer plugs him in, uses it to surf the web and do his job, and has had new features added over time. The account is a rare longitudinal look at sustained, real-world use of a speech-decoding BCI rather than a one-off demonstration.

This work studies how to fuse the knowledge of multiple masked diffusion language models (MDLMs). The authors find successful generations show stable confidence dynamics over answer-relevant positions, while unreliable trajectories can often be corrected by injecting promising intermediate states from another model. TIE (Trajectory-based Iterative Ensembling) tracks confidence over answer-relevant positions to decide which model currently follows the more reliable trajectory and relays partially denoised states across models. It is an early ensembling recipe tailored to the distinct decoding dynamics of diffusion LMs rather than autoregressive ones.

Multi-turn serving accumulates KV cache that grows with every turn and user, making memory the binding constraint on throughput. Non-uniform compression, allocating different budgets per attention head, preserves accuracy better than uniform schemes but breaks serving stacks that assume identical KV lengths, trapping freed memory as page fragmentation and inflating decode latency. Tangram exploits a two-level structural regularity, an input-invariant head ranking with narrowly bounded variation, so head-wise retention can be planned ahead rather than discovered at runtime, recovering the accuracy benefit without the serving-time penalties.

The Arbiter monitors multi-agent conversations in real time to identify which participants may be behaving in misaligned ways, addressing the case where individually well-aligned agents produce problems through their interaction. Operating under a limited inspection budget, it observes step by step and chooses to wait, question a participant, examine internal information such as system prompts or reasoning traces, or log concerning behavior, finally producing a report on the likely source of misalignment. It targets oversight of the multi-agent systems increasingly used to negotiate and act on shared tasks.

This paper studies a subtler failure of AI-generated peer review than prompt injection: no hidden text and no change to methods, experiments, figures, equations, or results, only presentation-level edits to the abstract, contribution framing, related work, and narrative. The proposed 'adversarial repackaging' is a closed loop that uses AI-reviewer feedback to search for presentation revisions while holding the science fixed. Across three mainstream AI reviewers it reaches a 75.1% attack success rate with a mean score gain of +1.21 out of 10, an effect the authors show is not explained by ordinary prose polishing, raising direct questions for AI-assisted review pipelines.

Nvidia said it plans to raise 25 billion dollars in new debt, its first corporate bond sale since 2021 (when it raised 5 billion), even as it generates tens of billions in cash each quarter, per The Information. The move follows other large technology and AI-infrastructure players turning to debt markets to finance the capital intensity of the buildout, and signals that even the most cash-rich chipmaker sees value in locking in bond financing for AI-era expansion.

TechCrunch reports that in April, for the first time, an Earth-observation satellite identified what it was looking for entirely on its own, onboard, rather than downlinking raw imagery for ground processing. Autonomous on-orbit detection compresses the sense-to-decision loop and reduces bandwidth needs, with implications for both commercial remote sensing and defense ISR, where tasking and reaction times matter and where edge autonomy in space is an active area of investment.

BRDFusion unifies inverse and forward rendering of urban scenes from captured video by pairing complementary models: a physically-based model recovers explicit, consistent scene properties and controls lighting physics, while a generative model supplies priors that resolve optimization ambiguity and denoises away the reconstruction artifacts physical methods leave behind. The result targets high-quality, controllable video with applications in content creation and autonomous-driving simulation, where both physical fidelity and editability matter. Surfaced across three sources with 16 HF upvotes.

NewCore emerged with 66 million dollars in funding on the thesis that the next enterprise-security challenge is managing AI agents rather than human users, per TechCrunch. As agents take on employee-like roles with access to systems and data, the company is building identity and access infrastructure for non-human actors, an emerging category as organizations grapple with authentication, permissions, and auditability for autonomous software acting on their behalf.

A Washington, D.C. customer filed a class-action suit against Anthropic on Sunday night alleging it misled customers about the value of its premium Max 5x and Max 20x subscription plans, per The Information. The complaint centers on how the usage limits of the higher-tier plans were marketed relative to what subscribers actually received, part of broader scrutiny of how AI vendors communicate rate limits and usage caps on paid consumer plans.

Pythagoras-Prover is an open family of Lean theorem provers designed for practical compute budgets, spanning autoregressive models at 4B and 32B parameters plus a proof-of-concept 4B diffusion prover that iteratively refines Lean proofs at inference time. To control training cost it builds a Lean-verified corpus stratified into easy, medium, and hard problems for curriculum SFT, with a dynamic proof-reasoning filter that preserves informative traces, letting models acquire proof skills progressively. It targets the expensive SFT-and-sampling regime that has limited access to strong formal-math models.

mu_0 forecasts smooth 3D trajectories for salient interaction points, objects, tools, hands, and contact regions, rather than predicting dense pixels (which wastes capacity on appearance) or direct actions (which require embodiment-specific labels). This yields a compact, embodiment-agnostic motion interface that enables scalable robot learning. A TraceExtract system automatically derives 3D supervision from diverse video by selecting keypoints, building globally aligned traces, and associating motion segments with hierarchical language captions, allowing training from heterogeneous video sources.

World Tracing is a generative geometry representation that predicts 3D points aligned to observed pixels while also completing geometry beyond the visible surface, bridging depth estimators (anchored to pixels but stopping at the visible surface) and image-to-3D models (complete but often misaligned). For each pixel it predicts an ordered stack of camera-space 3D points, the first layer the visible surface and later layers front-to-back intersections with occluded surfaces, instantiated by a world-tracing diffusion transformer, WT-DiT, trained with pixel-space flow matching.

VisualClaw addresses three deployment gaps in VLM agents: high latency and cost on dense video frames and long prompts, static scaffolds that never improve after deployment, and benchmarks that don't test visual evidence use inside tool-using workspaces. It uses hybrid encoding, a cascaded gate that filters less-informative streaming frames plus hot/cold top-k compression of a text skill bank, and skill evolution, where retrieved memories condition an evolver to update the skill bank from past failures. Evaluated across four video-QA settings.

Bengaluru-based Sarvam reached unicorn status with a 234 million dollar round led by IT-services firm HCLTech, which is putting in 150 million dollars, per TechCrunch. Sarvam builds models and AI products oriented toward Indian languages and use cases, and the HCLTech-anchored raise reflects both domestic strategic interest in sovereign-leaning AI capability and the continued flow of large checks into national-champion model developers outside the US and China.

Meta announced a wave of new AI features on Facebook, including an 'AI Mode' that draws on public information across its platforms, part of its effort to catch up in the AI race and deepen engagement, per TechCrunch. The cross-platform sourcing of public data to ground the assistant's responses is the notable detail, extending Meta's consumer AI surface deeper into its largest social product.

C4ISRNET reports that battlefield demand from Ukraine for compact laser-targeting systems on small drones has driven a vendor to launch a new product aimed at that niche. The item reflects the continued, rapid commercial response to attritable-drone warfare, where lightweight targeting and guidance payloads for inexpensive uncrewed systems have become a fast-moving procurement category.

Cohere announced a new London office, framed as roughly tripling its UK footprint and placing the company within London's AI research ecosystem as it grows a global research hub. The expansion is modest news on its own but fits Cohere's enterprise- and sovereignty-oriented positioning and the broader pattern of frontier labs anchoring research presence in the UK.