Wolf Digest — Monday, May 25, 2026

DiTs inherited the original Transformer's residual stream unchanged, while every other axis (tokenization, attention, conditioning, objectives, latent autoencoders) has been re-examined. This paper diagnoses three symptoms of vanilla residual addition in DiTs — monotonic forward magnitude inflation, sharp backward gradient decay, and pronounced block-wise redundancy — across both depth and denoising timestep. Their drop-in fix, Diffusion-Adaptive Routing (DAR), performs learnable, timestep-adaptive, non-incremental aggregation over the history of sublayer outputs and is compatible with REPA. On ImageNet 256×256 it improves SiT-XL/2 by 2.11 FID (7.56 vs. 9.67) and matches the baseline's converged quality with 8.75× fewer training iterations. Stacked on REPA it yields a 2× training acceleration in the early stage. 79 HF upvotes.

At AUSA's LANPAC and the Indo-Pacific Security Forum in Hawaii, INDOPACOM and Army logistics leadership framed sustainment, not posture, as the binding constraint on a Pacific fight. Gen. Brunson (US Forces Korea) said directly that "we cannot win if our supply lines are 5,000 miles long"; the geometry is Hawaii at 3,000 mi from CONUS, Guam at 5,000 mi from Hawaii, and the first island chain another 1,500 mi from Guam. Maj. Gen. Gavin Gardner (8th Theater Sustainment Cmd) said the Army is expanding "fix forward" contracts in Korea, Japan, the Philippines, Australia, and Singapore to eliminate 30-day tows back to CONUS for broken watercraft (a Talisman Sabre pain point two years running); Korean dry docks have completed 3 US ship overhauls with 2 more queued. INDOPACOM strategy chief Maj. Gen. Rowell cited China's >50% share of global commercial shipbuilding capacity vs. ~0.1% for the US and called resilience a warfighting function. Marines plan to self-sustain 45 days inside the first island chain but cannot pre-stage an "iron mountain."

At SOF Week in Tampa, SOCOM officials said operators are already using generative AI "heavily" for resource allocation and force deployment but need models that run at the tactical edge in disconnected environments — not just in cloud-connected data centers. The command is exploring fog-computing architectures to push cloud-class compute closer to collection and smaller LLMs that retain instruction-following with less compute. Acquisition exec Melissa Johnson said the relevant capabilities will likely come from smaller startups, since these niche tactical workloads aren't on hyperscaler roadmaps. Program-manager priorities: voice-command interfaces to reduce operator cognitive load (Col. Robert Oliver, PEO SOF Warrior), drone-to-drone interoperability and gesture/voice mission planning (Lt. Col. Aaron Davidson, unmanned autonomy), and agentic AI that can plan, revise, and execute (Rob McClintock, intel PM).

StepAudio 2.5 is a unified audio-language foundation model whose central claim is that once text and audio share a multimodal representational space, the gap between unified models and specialists collapses to operational regime — data construction, optimization targets, and decoding constraints — rather than architecture. The authors advance the post-training pipeline from standard supervised learning to task-tailored RLHF, shaping a shared backbone into three operational modes: an ASR branch using verifiable multi-token decoding for transcription efficiency, a TTS branch using preference-based RLHF and context-rich supervision for controllable expressive synthesis, and a realtime spoken-interaction branch optimized for low latency. The report positions StepAudio 2.5 as matching or exceeding specialized systems on all three. 34 HF upvotes.

A 22-page Mitchell Institute paper by retired Space Force Col. Kyle Pumroy argues USSF must build its own human-spaceflight program and expand Title 10 active-duty authorities to cover "space and lunar habitation" and warfighting tasks, citing China's stated 2030 crewed lunar goal. The report explicitly recommends a strategic vision "unconstrained by" the 1967 Outer Space Treaty (which bans military bases and maneuvers on celestial bodies), and urges Congress to fund commercial space-station residencies and potentially a dedicated USSF station in future NDAAs. USSF has loaned officers to NASA (Mike Hopkins transferred in-orbit in 2020; Col. Nick Hague commanded SpaceX Crew-9 for 171 days in 2024) but has no operational human spaceflight today. Secure World Foundation's Victoria Samson noted the report reflects a deliberate blurring of the longstanding exploration / militarization separation under an expansionist Space Force posture.

Google Cloud COO Francis de Souza argues AI security must be built into the platform from the start and warns that "shadow AI" will surface forgotten data assets (old SharePoint servers) whose access controls were never updated, and cites that mean time from initial breach to next attack stage has compressed from 8 hours to 22 seconds — forcing "AI-native, fully agentic defense" with humans overseeing rather than in-the-loop. TechCrunch pairs the framing with The Register's reporting on Google Cloud devs hit with five-figure surprise bills from unauthorized Gemini API calls: Maps API keys, deployed publicly per Google's own instructions, were silently scoped to access Gemini after a policy expansion. Prentus's Rod Danan was billed $10,138 in ~30 min; Sydney dev Isuru Fonseka was charged ~AUD $17,000 despite a $250 cap, because Google's automatic tier upgrades had raised ceilings to as high as $100,000 without consent. Aikido researcher Joseph Leon found revoked API keys remain usable for up to 23 minutes (success rates >90% in some minutes) due to slow revocation propagation, while service-account creds revoke in ~5 sec and AQ-prefixed Gemini keys in ~1 min — indicating the gap is policy, not engineering. Google has no plans to change the auto-tier policy.

SciAtlas argues current academic retrieval tools rely on keyword matching or vector-space semantic retrieval that lacks the topological reasoning to navigate complex logical connections, and that agentic deep-research frameworks are prone to logical hallucinations and high inference costs. The authors build a large-scale knowledge graph over the scientific literature to support automated scientific research, intended as a substrate that combines structural reasoning with vector retrieval. 39 HF upvotes.

Companion-paper energy to SkillOpt: this work runs the first comprehensive study of model-generated agent skills across the full lifecycle — experience generation, skill extraction, and skill consumption — to ask whether such skills actually act as durable procedural artifacts. Domain-level and model-generated skills are flagged as the most promising regime because they encode recurring procedures within a domain and scale beyond hand-crafting, but the empirical picture so far has been inconsistent. The paper systematizes the variables (which experiences to generate, which extraction method to use, which consumption harness) and reports findings across that grid. 19 HF upvotes.

ETCHR pursues a decoupled approach to the "think with images" paradigm: rather than relying on fixed toolkits or unified multimodal generation (which produces noisy intermediate images), use a dedicated image editing model alongside the understanding model. The paper diagnoses two failure modes of off-the-shelf image editors when used as reasoning assistants, and proposes targeted training to fix them. Targets fine-grained focus and view-transformation questions that purely-textual chains of thought handle poorly. 9 HF upvotes.

SWIM aligns vision and language representations to enable fine-grained object understanding from textual prompts alone, without requiring explicit visual prompts like masks or points at inference. Mask supervision is used only at training time to guide cross-modal attention so the model learns to attend to user-specified objects automatically. Cross-attention analysis of pretrained MLLMs reveals a systematic discrepancy: attribute words produce sharp, localized attention while object nouns do not — the paper's training strategy closes that gap. 28 HF upvotes.

PhotoFlow tackles virtual photography in prepared 3D scenes: given scene information and a language intent (no preselected camera pose or reference image), the agent must infer a shot, choose executable camera parameters, and render the photograph. The closed-loop architecture has three roles — a Director that proposes camera setups, a Reviewer that scores them, and a Reflector that iterates — stressing 3D spatial understanding and abstract aesthetic judgment together. 20 HF upvotes.

VGenST-Bench critiques existing spatio-temporal reasoning benchmarks for relying on static image sets or passively curated video, which underconstrain the evaluation. Instead, the authors actively synthesize highly controlled and diverse evaluation scenarios using generative video models, enabling fine-grained probing of MLLM spatio-temporal capabilities. The multi-stage construction pipeline is the paper's main methodological contribution. 17 HF upvotes.

PiD attacks the megapixel-scale decoder bottleneck in latent-diffusion and autoregressive image systems. Reconstruction-oriented decoders, optimized to invert the encoder, become increasingly costly at high resolution and don't synthesize meaningful detail. PiD reformulates the latent-to-pixel step as pixel-space diffusion, leveraging recent advances in scalable pixel-space diffusion to provide a more expressive and efficient decoding paradigm. 13 HF upvotes.

Current post-training pipelines for discrete autoregressive T2I (VQ tokenizer + AR policy) only optimize the policy and keep the VQ decoder frozen. RankE shows this induces Latent Covariate Shift: as the policy evolves, the token distribution diverges from the ground-truth distribution the decoder was trained on, degrading reward scores. The paper proposes co-evolving the decoder with the policy, end-to-end. Analogous in spirit to the REPA-E line of work for diffusion T2I but for the discrete-AR family. 13 HF upvotes.

This paper argues recent emphasis on long chain-of-thought reasoning in VLMs misdiagnoses the bottleneck: performance on visual tasks is primarily limited by lack of visual perception, not reasoning itself. The authors decompose VLM capabilities into three training stages — visual perception, visual reasoning, and textual reasoning — and show visual perception requires targeted optimization with specialized data that won't emerge from generic post-training. 4 HF upvotes.

Long-horizon LLM-agent RL suffers from sparse outcome rewards that don't tell you which intermediate actions caused success or failure. Prior work generates per-turn feedback or hints, but feedback at every turn is wasteful when many turns are already neutral. HINT-SD targets hindsight self-distillation at the specific steps that actually mattered for the trajectory's outcome, improving sample efficiency over uniform feedback application. 3 HF upvotes.

Muon's Newton-Schulz spectral gradient orthogonalization (driving all singular values toward 1) wins on LLM pretraining but the paper identifies two failure regimes beyond pretraining: (i) cross-modality vision-language-action training, where low-rank action-module gradients cause amplification of noisy tail directions; (ii) reinforcement learning with verifiable rewards, where the same uniform whitening misallocates capacity. The authors propose high-pass remedies that preserve Muon's benefits in pretraining while correcting these failure modes. 4 HF upvotes.

Unified multimodal models often exhibit functional inconsistency between understanding and generation despite sharing a latent space. The authors observe this stems not from missing shared representations but from absence of explicit alignment between the transformations that map into and out of the latent space — generation and re-encoding follow inconsistent trajectories, causing semantic drift under modality transitions. LatentUMM enforces dual-latent alignment to close the inconsistency. 4 HF upvotes.

Existing scaling laws — predominantly monotonic power laws — fail to explain emerging non-monotonic phenomena like catastrophic overtraining and quantization-induced degradation, where performance worsens with more compute. The authors propose a Shannon Scaling Law grounded in Shannon-Hartley: model parameters as channel bandwidth, training tokens as signal power. The formulation explicitly captures how those interact and predicts the non-monotonic regions. 7 HF upvotes.

Interactive world models for FPS games must resolve high-frequency overlapping control signals every frame without disrupting unaffected regions. SCOPE observes that FPS actions are spatially selective: discrete events (firing, reloading) affect only a localized region around the weapon, while continuous camera/movement signals govern stable surroundings. They insert a conditioning module into each transformer block of a pretrained world model that respects this decomposition, enabling cross-game training. 8 HF upvotes.

Camera-controlled video generation has progressed quickly, but existing video-to-video re-rendering methods rely on SFT over synthetic datasets and there is an acute scarcity of synchronized multi-view real-world video. The result is limited generalization on out-of-distribution real videos, with models failing to adhere to physical scales and camera trajectories. Geo-Align is presented as the first RL framework that uses metric-geometry rewards to align generated video to true camera trajectories. 5 HF upvotes.

The paper challenges the under-examined assumption that current Vision-Language Models faithfully synthesize multimodal data. Their argument: state-of-the-art models in the Vision Encoder-Projector-LLM paradigm frequently exhibit functional blindness — exploiting strong language priors to bypass severe visual representation bottlenecks rather than extracting grounded visual knowledge. They frame this as a trustworthiness problem in the dominant monolithic VLM paradigm. 2 HF upvotes.

GenRecon casts high-fidelity 3D scene reconstruction from multi-view RGB images as conditional 3D generation over spatially-localized overlapping chunks that tile the scene, inheriting the fidelity and completeness of state-of-the-art generative shape models (Trellis.2 in their experiments) generalized to scene scale. A projection-based conditioning mechanism lifts multi-view images into the per-chunk generative prior. 2 HF upvotes.

Visual geometry transformers grow quadratically in compute with input sequence length because of global attention. This paper restricts the number of key/value tokens each query interacts with, introducing a two-stage selection strategy to keep the most informative tokens. The result preserves multi-view 3D reconstruction quality while enabling much longer input sequences and better scalability. 2 HF upvotes.

The Information frames the SpaceX mega-IPO and OpenAI's prep work as anchors of a growing 2026 listings calendar that now also includes smart-ring maker Oura and crypto company Blockchain.com. The piece uses recent IPO history to assess what signals investors should read into pricing and aftermarket performance, particularly relevant given the scale of the SpaceX float and the fact that an OpenAI listing would be the most-watched AI debut to date. Body of analysis is paywalled below the lede.

Tim De Chant reads the SpaceX S-1 as evidence Musk has pivoted from Tesla's "solar electric economy" vision: xAI is running its data centers on dozens of unregulated natural-gas turbines with plans to spend another $2.8B on gas generation, while SpaceX has purchased $131M of Cybertrucks (1,279 units) and xAI $697M of Tesla Megapacks over two years — but materially zero Tesla solar panels. The filing instead pitches space-based solar, claiming orbital arrays generate >5× terrestrial output thanks to 24/7 illumination, and projects "terawatt-scale annual AI compute growth" against ~40 GW of current global data-center load. De Chant flags economics — Starlink power costs are multiples of terrestrial DC power, radiation-hardening chips is expensive, and it's unclear AI training can be sharded across many satellites — making the orbital-DC thesis a high-variance bet currently backstopped by fossil fuels.

Lucas Ropek's hands-on with Bee, the wrist wearable Amazon acquired last year: a button-toggled always-listening recorder that pushes audio to a cloud-backed app for transcription, summaries, and calendar-linked reminders. Summaries were "moderately competent" on business calls and correctly labeled a Reservoir Dogs viewing as "Tarantino Film Scene Analysis" — but transcripts miss segments and don't reliably attribute speakers without manual labeling, performance not materially differentiated from Otter or Granola. The reservation is permissions surface (location, photos, contacts, calendar, notifications, optional health data) and cloud storage; Bee reportedly demoed a fully local-running version to a YouTuber but Amazon has given no roadmap. Stated security posture: encryption at rest/in-transit, third-party audits, continuous monitoring.

Ferrari's IBM-powered fan-app rebuild is IBM's first major F1 partnership, chosen over rivals after Anthropic, AWS, and Oracle each moved into the sport. The Scuderia Ferrari HP app adds AI-written race summaries, an AI Q&A companion, in-app games, predictions, and Italian-language support for the first time. IBM cites a 62% jump in engagement over race weekends since launch; Ferrari is using AI to analyze in-app engagement signals and fan-message sentiment to drive content personalization. Context: each F1 car generates millions of telemetry data points per second per race, and 75% of new F1 fans last year were women — many Gen Z — accelerating direct-to-fan strategies.

A Memorial Day reflection by Army officer Adam A. Scher (now at JIATF-401, the Pentagon's counter-drone office) profiling three fallen soldiers — Cpl. Andrew J. Kemple (KIA Tikrit 2006), 2nd Lt. Tracy Lynn Alger (KIA Iraq IED 2007), Sgt. David Scott Robinson (Highway 1, Afghanistan, 2010) — and drawing operational-leadership lessons from each: the cost of grief on mission continuity, the protest-at-funerals fight Kemple's family triggered into NY state law, and the COIN-era tradeoff between civilian disruption and convoy risk. Not a policy piece — a remembrance essay aimed at command leaders.

The Information's weekend newsletter package, with the lead Big Read profiling Stanford ex-president Marc Tessier-Lavigne's post-resignation trajectory and a separate piece on the "tech elite's go-to real estate broker." Most of the body is paywalled. Surfaced here because it appeared in The Information's AI feed in-window and includes the crypto-Washington and books pieces also indexed by the feed.

The Information profiles 1155 F Street in Washington, a 12-story glass-paneled office tower that has become the operational hub of crypto's Washington influence operation. The piece walks through which firms have moved in and how the industry is coordinating its lobbying push for regulatory clarity. Paywalled below the lede. Tangentially AI-adjacent via the broader DC-tech footprint pattern.