AI Digest

Category

Nvidia's board orders an independent review after Jensen Huang admits an affair amid spy claims

The Rundown AI

Nvidia supplies nearly the entire AI industry, so questions about its CEO could have effects well beyond the company.

  • Nvidia's board ordered an independent review of reports that Huang was targeted by a "foreign intelligence effort". In a letter to staff, Huang admitted to an extramarital affair.
  • A spring memo reportedly linked Alice Huang, an Nvidia customer executive, to China's top spy agency. China denies the claims.
  • Huang remains CEO. The board says it sees "no reason to believe" any secrets were compromised, and Huang has until Nov. 11 to answer a House GOP request for a briefing.
Read on nvidianews.nvidia.com →

OpenAI reportedly raises $11B at a $1.25 trillion valuation

TLDR AI

OpenAI's valuation keeps rising even as investors grow less sure about AI valuations.

  • The round could reach $20B before it closes. That is still short of the $30B target reported earlier.
  • The valuation is up from $852B in March.
Read on techcrunch.com →

An AI research system using an Anthropic model proves the Aaronson-Ambainis conjecture

TLDR AI

An AI system has resolved a major open problem in quantum complexity theory.

  • Physical Superintelligence (PSI) built the research system, which ran on an unreleased Anthropic model. The conjecture had been open since 2008.
  • The proof shows that quantum computers are not exponentially faster than classical ones on random inputs.
  • Stanford researchers called it "a huge result."
Read on psi.inc →

Microsoft Research's AI Frontiers group has been folded into Microsoft AI

TLDR AI

The change moves Microsoft's research on making agents work into its product-focused AI division.

  • AI Frontiers was a sub-organization within Microsoft Research that coordinated research efforts, much of them focused on making agents work.
  • It was reorganized into Microsoft AI in early September.
Read on hunch.net →

OpenRouter launches a new Auto Router and Auto Code, a router for coding work

TLDR AI

Automatic model routing gets better while staying within each team's governance rules.

  • OpenRouter says the new Auto Router outperforms the previous one by up to 4x.
  • Both routers weigh eleven factors about each request, plus live pricing, provider health and cache status.
  • Users can apply their own governance policies, and the router picks from the models those policies allow.
Read on openrouter.ai →

Meta adds Spaces to its Muse app for building custom interactive experiences

TLDR AI

Meta's personal agent app now lets people build small apps by describing them.

  • Users describe what they want, such as a gift tracker, a game or a project planner, and Muse builds it.
  • Muse also now includes talking avatars.
Read on about.fb.com →

Higgsfield previews Miracle, which splits video into editable layers

The Rundown AI

Layer-based editing gives users more precise control over AI video.

  • Miracle lets users recolor objects or swap on-screen text by editing separate layers of a video.
  • It is currently open only to beta testers.
Read on x.com →

NVIDIA and Elith release Nemotron-Personas-Finance-Japan, 9M synthetic financial personas

Hugging Face Blog

Japanese banks and insurers can build and test financial AI without using real customer data.

  • The dataset has 1M records with 9 personas each and 50 fields, grounded in official Japanese statistics on demographics, household finances and financial literacy.
  • It was built with NeMo Data Designer and gemma-4-31B-it, and is released under CC BY 4.0.
  • Nemotron-PII-Japan, with 100,000 synthetic business documents labeled for personal information, launched the same day.
Read on huggingface.co →

Ace lets developers and agents build software together, locally

TLDR AI

Teams can collaborate with agents without exposing anything to the public internet.

  • Developers work with agents and teammates from their own machines.
  • Agents run with the privileges of the user running the host, so they can execute shell commands and change files outside the project.
Read on ace.githubnext.com →

ByteDance Seed uses sparse autoencoders to map what individual MoE experts do

Hugging Face Daily Papers

Shows that MoE experts specialize at a finer level than coarse domains, which helps explain how these widely used models divide work.

  • Trains SAEs on MoE layer outputs so each latent's pre-activation decomposes exactly across experts, measuring contribution rather than routing frequency.
  • On layers 4, 30 and 46 of Qwen3-30B-A3B, contribution is far more concentrated than gating, and most latents have only one or two dominant experts.
  • Finds distinct experts for Java vs. front-end web code and for geometry vs. arithmetic and polynomials; SAEs and code are released.
Read on huggingface.co →

StepFun's StepX-OPD merges multi-teacher distillation and few-step sampling into one stage

Hugging Face Daily Papers

Shows a way to build deployable few-step image models without losing quality in a separate compression stage.

  • One student learns several capabilities from frozen specialist teachers while it learns to sample in a few steps, with supervision on its own rollout states.
  • On SD3.5-Medium and FLUX.1-dev, it gets the highest normalized average among the four-step methods compared and beats the forty-step Flow-OPD model.
  • Trains with about 37% fewer GPU-hours than sequential DiffusionOPD followed by step distillation on SD3.5-Medium.
Read on huggingface.co →

SwarmEvolve runs evolutionary program search with agent swarms that reason in latent space

Hugging Face Daily Papers

Makes AlphaEvolve-style discovery much faster and cheaper by skipping text chain-of-thought.

  • Hundreds of low-cost agents per round share information through KV caches, and a novelty filter keeps ideas whose hidden states are farthest apart.
  • On 21 math and scientific-computing problems, it reaches the best scores that other frameworks reach in 60 minutes in a median of 3 to 8 minutes.
  • With Qwen3 8B, it improves on several solutions that AlphaEvolve first found using frontier LLMs.
Read on huggingface.co →

Snowflake's KernelSeek uses targeted questions, not extra tools, to guide CUDA kernel optimization

Hugging Face Daily Papers

Suggests coding agents need direction more than more tooling to tackle hard performance work.

  • On the same task, a tool-heavy autonomous agent plateaued at 0.50x FlashAttention-3 while KernelSeek reached 1.13x.
  • Produced dependency-free FlashAttention forward kernels for Hopper and Blackwell at 1.16x FlashAttention-3 and 1.46x FlashAttention-4 (geometric mean), correct on all 182 scenarios per architecture.
  • Releases the kernels, harness, question bank and all traces.
Read on huggingface.co →

Multi-layer Indexer speeds up DeepSeek Sparse Attention decoding without retraining

Hugging Face Daily Papers

Adds a free speedup for long-context serving of DeepSeek-V3.2 and GLM-5.

  • Since adjacent layers pick nearly the same tokens, each layer attends to the previous indexer's selection, so indexer scans can run at the same time as attention.
  • Combined with IndexCache, it decodes up to 14.5% faster on DeepSeek-V3.2 and 9.8% faster on GLM-5 at retention 1/2.
  • Reaches 1.32x the unmodified model's decode throughput at 128K context, inside SGLang's production decode graph.
Read on huggingface.co →

Identically trained language models fail on different jailbreaks depending on pretraining seed

Hugging Face Daily Papers

Shows that jailbreak benchmark results on a single checkpoint may not carry over to retrained models.

  • Across 150 1B models, per-prompt non-refusal rates correlate at 0.984 between post-training seeds but drop to 0.811 with a different pretraining seed.
  • The best attack template on one checkpoint stays the best after a new pretraining seed in only 27% of model pairs.
  • Per-prompt refusal patterns identify a model's pretraining run among all 75 runs every time.
Read on huggingface.co →

OpenFrontiers benchmarks whether LLM agents can rediscover recently solved math results

Hugging Face Daily Papers

Tests research-level problem solving in a way that separates real reasoning from looking up published answers.

  • Pairs a corpus of 12,151 open research problems with a 200-problem benchmark stated in Lean 4, including 110 problems resolved in 2026.
  • Across seven agents, closed-book rediscovery is rare: at most 7 of 40 resolved problems.
  • With live web access, success mostly comes from finding published solutions.
Read on huggingface.co →

Coca-Cola compresses agent tool outputs into latent tokens instead of text

Hugging Face Daily Papers

Cuts agent context costs sharply while improving task results.

  • Compresses tool observations into latent tokens in the agent's input embedding space, trained so the agent picks the same next actions as with full observations.
  • Reduces token use by 70.01%–89.24% and end-to-end runtime by 24.45%–55.97% on four web-search QA benchmarks.
  • Improves task performance consistently across two model families; code is on GitHub.
Read on huggingface.co →

HeLM gives robot VLA models explicit language memory for tasks that depend on history

Hugging Face Daily Papers

Takes on a key limitation of robot policies that act only on the current observation.

  • Keeps task-relevant history as language and updates it only when it detects a relevant event.
  • Improves within-episode success by 12.8 percentage points over the strongest baseline across 10 real-world tasks with 56 variants.
  • Uses up to 4.5x less memory and runs 2.5x faster than a keyframe baseline at long horizons.
Read on huggingface.co →

NitroFusion2 makes one-step image generation practical with adversarial loss on the text stream

Hugging Face Daily Papers

Cuts few-step diffusion distillation from 4–8 steps to 1–2 at comparable quality.

  • Uses intermediate text-token states of multimodal diffusion transformers as the representation for the adversarial discriminator.
  • Conditions the discriminator on aspects such as typography, objects or style to steer what training emphasizes.
  • At one step, it produces sharp rendered text and fine detail.
Read on huggingface.co →

Anthropic explains how Claude agents "dream" to reorganize their memory

The Rundown AI

A frontier lab's account of how agents can keep improving over long-running work.

  • During downtime, Claude checks the notes it has saved as files against logs of past work to catch errors.
  • Anthropic found that explaining why memory matters works better than giving rules about what to write.
  • In Pokémon tests, Fable 5 earned four badges with 10 memory files, Sonnet 3.7 earned one with 413, and Opus 5.5 beat the game.
Read on anthropic.com →

Project EVE: Anthropic tests Claude's decision-making in MMORPGs

TLDR AI

Large multiplayer games are a demanding testbed for complex agent decisions.

  • Anthropic used Claude to test complex decision-making in EVE Online and EVE Frontier.
Read on anthropic.com →

OpenAI introduces BioBasicsBench for routine wet-lab science

TLDR AI

The benchmark tests whether AI agents can interpret raw lab data from everyday experiments.

  • It evaluates how well AI interprets raw experimental data from 100 routine wet-lab tasks.
Read on openai.com →

ByteDance Seed's Hierarchical Residuals separate residual width from compute width

Hugging Face Daily Papers

A simple architecture change that improves FLOPs efficiency using only standard operators.

  • HiRes widens the residual state between blocks of layers while keeping each layer's compute width unchanged.
  • Fitted scaling laws show a 1.30x gain in FLOPs efficiency.
  • On OLMoE-1B-7B trained for 500B tokens, training loss is 0.0208 lower and downstream accuracy is 1.14 points higher.
Read on huggingface.co →

Amazon's HOPE prunes MoE experts while keeping groups of cooperating experts together

Hugging Face Blog

Removing half of a mixture-of-experts model's experts no longer has to hurt agentic coding performance.

  • HOPE is a second-order method that solves a quadratic program over pairwise interactions between experts.
  • At 50% pruning, Qwen3.5-35B-A3B scores 42.3 on SWE-Bench Pro, against 41.8 for the unpruned model and 36.2 with REAP.
  • Code is available on GitHub and has been tested on Qwen3, Qwen3-Next, Qwen3.5 and GLM-4.5.
Read on arxiv.org →

NVIDIA's AgentBaton hands long agent tasks off from expensive models to cheaper ones

Hugging Face Daily Papers

A well-timed handoff to a cheaper model can cut agent costs sharply without losing accuracy.

  • Handing off from capable to efficient models at the right stage largely preserves accuracy. Going the other way underperforms.
  • Across 11 model pairs, AgentBaton cuts API costs by an average of 65.9%, with accuracy gains of up to 8.8 points.
Read on huggingface.co →

NVIDIA's AutoCompress prunes models to fit hardware constraints, not parameter counts

Hugging Face Daily Papers

Compression aims directly at real latency, throughput and memory limits.

  • It builds serving-engine cost models into differentiable architecture search.
  • It beats Minitron's average accuracy by 6.57, 2.35 and 0.57 points on RTX 5070, RTX 5080 and H100.
  • It compresses Nemotron 3.5 Super from 120B to 74B parameters for single-B200 serving at 128K context, keeping 95% of its benchmark score.
Read on huggingface.co →

HEELO builds agent harnesses from reusable "agent-native primitives"

Hugging Face Daily Papers

Frontier agents still fail many real-world tasks, and this framework more than quadruples completions on the authors' platform.

  • Dev-Primitives raise Terminal-Bench 4.0 resolution from 37.3% to 55.5% on the same Codex backbone.
  • On 46 real-world tasks, Claude Opus 5.5 completes 39 with HEELO versus 9 alone, and GPT-6.1 Sol completes 35 versus 9.
Read on huggingface.co →

IWIBench measures how well models can edit another model's weights directly

Hugging Face Daily Papers

Editing a model's weights without gradients is emerging as a new capability of frontier models.

  • Editor models patch a smaller target model, usually Gemma 3 4B IT, to meet ten text-editing objectives.
  • Astra and GPT-6.1 Sol lead by a wide margin, while GPT-5.6 Sol shows little of this ability.
  • Structured-response tasks resisted every attempt.
Read on huggingface.co →

StepFun's P^2-MOPD speeds up multi-teacher distillation by using expert weights as priors

Hugging Face Daily Papers

Consolidating expert models into one student becomes faster and more balanced.

  • It jointly optimizes the mixing coefficients for student and expert weights along with the student's own weights.
  • The 1.7B model reaches 94.7% average recovery, versus 80.9% for MOPD, and passes MOPD's final result in half the training steps.
Read on huggingface.co →

Nothing here yet.