Jensen Huang resigns as Nvidia CEO; CFO Colette Kress named interim CEO
Nvidia, the company that supplies most of the hardware behind AI, has lost its founder and has no permanent successor yet.
- Huang resigned as CEO and board member, effective immediately, after reports of extramarital affairs, including one with an alleged Chinese intelligence asset.
- Kress, CFO since 2013, will hold both roles while the board considers internal and external candidates. Long-time director Harvey C. Jones becomes Nvidia's first board chair.
- Nvidia will still report third-quarter fiscal 2027 results on November 17 as scheduled.
Google releases Gemini 4 Hydrogen, a fast, low-cost model for coding and agents
Google's new efficient model scores higher than its Gemini 4 Argon model on a standard intelligence index and costs less to run.
- It scored 51 on the Artificial Analysis Intelligence Index, beating Gemini 4 Argon at its highest thinking level.
- Pricing is $1.50 per million input tokens and $7.50 per million output tokens.
- It has a 1M-token context window.
Benchmarks
| Benchmark | Gemini 4 Hydrogen | Claude Fable 5.2 | GPT-6.2 Astra | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.8 | 96.0 | 96.8 | — | 94.2 | +0.0 |
| FrontierMath-Tiers-1-3-v2-Private | 88.1 | 95.1 | 92.6 | 82.1 | 78.2 | -7.0 |
| FrontierMath-Tier-4-v2-Private | 87.8 | 100.0 | 100.0 | 70.7 | 65.9 | -12.2 |
| OTIS Mock AIME 2024-2025 | 100.0 | 99.7 | 100.0 | 99.4 | 100.0 | +0.0 |
| SimpleQA Verified | 70.9 | 94.0 | 78.5 | — | 62.2 | -23.1 |
| Chess Puzzles | 60.0 | 79.0 | 78.0 | — | 45.0 | -19.0 |
| EBR-bench | 70.5 | 90.5 | 92.4 | 31.4 | 37.1 | -21.9 |
| Mystery Game Puzzles | 65.0 | 90.0 | 95.0 | 51.0 | 42.0 | -30.0 |
| SimpleBench | 91.7 | 94.9 | 79.2 | — | — | -3.2 |
| APEX-Agents | 78.0 | 80.1 | 72.2 | 66.1 | 65.5 | -2.1 |
| Furniture Assembly | 57.5 | 90.0 | 89.2 | 54.2 | 55.8 | -32.5 |
| FrontierSWE | 53.1 | 69.9 | — | — | — | -16.8 |
| SpireClimb | 7.7 | 15.3 | 41.3 | — | 0.0 | -33.6 |
Δ = Gemini 4 Hydrogen minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Forensic analysis finds signs of export-controlled Nvidia Blackwell GPUs in top Chinese models
The analysis suggests several leading Chinese labs are training on chips that US export controls are meant to keep from them.
- Six models released since April by DeepSeek, Moonshot AI, MiniMax and Ant Group contain numerical patterns that match Blackwell hardware.
- Former OpenAI researcher Will DePue made the claim, citing GPU fingerprints found in the published model weights.
- The analysis does not show where the chips were, how access was obtained, or whether export controls were broken.
Nvidia's board is reportedly preparing to remove Jensen Huang as CEO after a second affair report
Removing the CEO of the company whose chips train nearly every major AI model would be a big shock to the AI hardware supply chain.
- The Wall Street Journal reports that directors are preparing to push Huang out, hours after The New York Times reported an affair with Nvidia health care VP Kimberly Powell, who reports to him.
- The Economist and members of Congress have called for him to step down over his involvement with alleged Chinese spy Alice Huang.
- Nvidia lost more than $427B in market value over two sessions after the first affair came out, and fell about 8% after hours as the new reports landed.
OpenAI reportedly posted its first $10B quarter
OpenAI's revenue nearly tripled in a year, a sign of how fast spending on frontier AI is growing.
- OpenAI reportedly made $10B in revenue in Q3.
- That is close to triple what it made a year earlier.
Google DeepMind releases AlphaGenome 2, which predicts how DNA variants affect specific cell types
Tying genetic variants to the cell types they affect helps researchers decide which disease experiments to run first.
- On a test across 28 immune cell types, it caught 30% of the DNA changes behind gene shifts, up from 16% for the original.
- It can take gene activity from just a few dozen cells and predict new cell states without retraining.
- It is available for non-commercial use.
Google releases Gemini 4 Carbon, its strongest model, to developers and enterprises
Google is claiming top spots on agentic coding and automation benchmarks at a lower cost per task than any other external model.
- Available through the API in AI Studio and Gemini for business, and rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search; priced at $2/$10 per million input/output tokens, the same as Gemini 4 Argon.
- Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best external model.
- The release adds frontier cybersecurity capabilities through the Fairwind Program and previews /swarm, an Antigravity mode where named agents split a task and coordinate on a shared message board.
Benchmarks
| Benchmark | Gemini 4 Carbon | Claude Fable 5.2 | GPT-6.2 Astra | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.5 | 96.0 | 96.8 | — | 94.2 | -0.3 |
| FrontierMath-Tiers-1-3-v2-Private | 90.5 | 95.1 | 92.6 | 82.1 | 78.2 | -4.6 |
| FrontierMath-Tier-4-v2-Private | 95.1 | 100.0 | 100.0 | 70.7 | 65.9 | -4.9 |
| OTIS Mock AIME 2024-2025 | 100.0 | 99.7 | 100.0 | 99.4 | 100.0 | +0.0 |
| SimpleQA Verified | 77.7 | 94.0 | 78.5 | — | 62.2 | -16.3 |
| Chess Puzzles | 78.0 | 79.0 | 78.0 | — | 45.0 | -1.0 |
| EBR-bench | 87.6 | 90.5 | 92.4 | 31.4 | 37.1 | -4.8 |
| Mystery Game Puzzles | 73.0 | 90.0 | 95.0 | 51.0 | 42.0 | -22.0 |
| ARC-AGI | 98.5 | 98.5 | 98.5 | — | — | +0.0 |
| SimpleBench | 92.6 | 94.9 | 79.2 | — | — | -2.3 |
| APEX-Agents | 84.2 | 80.1 | 72.2 | 66.1 | 65.5 | +4.1 |
| ARC-AGI-2 | 94.2 | 93.8 | 95.8 | — | — | -1.6 |
| PostTrainBench | 41.0 | 44.6 | — | 24.2 | — | -3.6 |
| LMCA | 65.7 | 70.6 | — | 56.8 | 57.6 | -4.9 |
| DTBench | 98.4 | 99.2 | — | 95.2 | 93.3 | -0.8 |
| Furniture Assembly | 77.5 | 90.0 | 89.2 | 54.2 | 55.8 | -12.5 |
| FrontierSWE | 61.2 | 69.9 | — | — | — | -8.7 |
Δ = Gemini 4 Carbon minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Anthropic launches its IPO roadshow, offering 825M shares at $70–$78
The company has moved from preparing for an IPO to setting a price range and a Nasdaq ticker.
- Anthropic plans to offer 825 million shares priced at $70–$78 each.
- It applied to list on Nasdaq under the ticker ANT, and most proceeds from newly issued shares will go to Anthropic.
- The Rundown reports the IPO is expected next week.
Anthropic says it now considers its AI R&D capability threshold met
A frontier lab is formally saying its models are reaching the point where they can meaningfully speed up AI research.
- The statement appeared in Anthropic's updated IPO filing.
- Anthropic based the judgment on what it expects from upcoming models.
Anthropic releases Claude Fable 5.2 and Mythos 5.2
Anthropic's new top models take first place on a major leaderboard while costing less and running faster than the previous version.
- Fable 5.2 debuted at No. 1 on Artificial Analysis' Intelligence Index, 5 points ahead of Google's unreleased Gemini 4 Carbon.
- Mythos is limited to vetted security and life sciences groups. Fable is the same model with stronger cyber and biology guardrails.
- 5.2 is 20% cheaper and up to 30% faster than 5.1, priced at $8 in / $40 out per million tokens.
Benchmarks
| Benchmark | Claude Fable 5.2 | GPT-6.2 Astra | Gemini 4 Carbon | Muse Wonder | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| GPQA diamond | 96.0 | 96.8 | 96.5 | — | 94.2 | -0.8 |
| FrontierMath-Tiers-1-3-v2-Private | 95.1 | 92.6 | 90.5 | 82.1 | 78.2 | +2.5 |
| FrontierMath-Tier-4-v2-Private | 100.0 | 100.0 | 95.1 | 70.7 | 65.9 | +0.0 |
| OTIS Mock AIME 2024-2025 | 99.7 | 100.0 | 100.0 | 99.4 | 100.0 | -0.3 |
| SimpleQA Verified | 94.0 | 78.5 | 77.7 | — | 62.2 | +15.5 |
| Chess Puzzles | 79.0 | 78.0 | 78.0 | — | 45.0 | +1.0 |
| EBR-bench | 90.5 | 92.4 | 87.6 | 31.4 | 37.1 | -1.9 |
| Mystery Game Puzzles | 90.0 | 95.0 | 73.0 | 51.0 | 42.0 | -5.0 |
| ARC-AGI | 98.5 | 98.5 | 98.5 | — | — | +0.0 |
| SimpleBench | 94.9 | 79.2 | 92.6 | — | — | +2.3 |
| APEX-Agents | 80.1 | 72.2 | 84.2 | 66.1 | 65.5 | -4.1 |
| ARC-AGI-2 | 93.8 | 95.8 | 94.2 | — | — | -2.0 |
| PostTrainBench | 44.6 | — | 41.0 | 24.2 | — | +3.6 |
| Remote Labor Index | 41.2 | — | — | — | — | — |
| ProofBench | 100.0 | — | — | 84.0 | 91.0 | +9.0 |
| LMCA | 70.6 | — | 65.7 | 56.8 | 57.6 | +4.9 |
| DTBench | 99.2 | — | 98.4 | 95.2 | 93.3 | +0.8 |
| Furniture Assembly | 90.0 | 89.2 | 77.5 | 54.2 | 55.8 | +0.8 |
| FrontierSWE | 69.9 | — | 61.2 | — | — | +8.7 |
| WeirdML v3 | 65.3 | — | — | 24.9 | 18.9 | +40.4 |
| SpireClimb | 15.3 | 41.3 | — | — | 0.0 | -26.0 |
Δ = Claude Fable 5.2 minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Meta releases Muse Ember, a small reasoning model for agent workloads
Meta now has a low-cost model aimed at the high-volume agent tasks that Anthropic and OpenAI's budget models handle.
- Ember tied Claude Haiku 5.5 at 36 on Artificial Analysis' Intelligence Index, well ahead of GPT-6 Luna at 28.
- An average test task cost $0.14 on Ember versus $0.41 on Haiku, because Ember uses fewer tokens per job.
- Meta trained Ember using its larger Muse Wonder Full, the first model trained that way.
Benchmarks
| Benchmark | Muse Wonder | Claude Fable 5.2 | GPT-6 Astra | Gemini 4 Argon | Kimi K3.1 Pro | Δ vs best other |
|---|---|---|---|---|---|---|
| FrontierMath-Tiers-1-3-v2-Private | 82.1 | 95.1 | 93.7 | 86.0 | 78.2 | -13.0 |
| FrontierMath-Tier-4-v2-Private | 70.7 | 100.0 | 97.6 | 78.0 | 65.9 | -29.3 |
| OTIS Mock AIME 2024-2025 | 99.4 | 99.7 | 100.0 | 100.0 | 100.0 | -0.6 |
| EBR-bench | 31.4 | 90.5 | 76.2 | 64.8 | — | -59.1 |
| Mystery Game Puzzles | 51.0 | 90.0 | 84.0 | 64.0 | 42.0 | -39.0 |
| APEX-Agents | 66.1 | — | 64.7 | 82.4 | 65.5 | -16.3 |
| ProofBench | 84.0 | 100.0 | 99.0 | 99.0 | 91.0 | -16.0 |
| LMCA | 56.8 | — | 64.4 | 63.7 | 57.6 | -7.6 |
| DTBench | 95.2 | — | 97.3 | 97.6 | 93.3 | -2.4 |
| Furniture Assembly | 54.2 | 90.0 | 80.0 | 72.5 | 55.8 | -35.8 |
| WeirdML v3 | 24.9 | — | 54.0 | 41.4 | 18.9 | -29.1 |
Δ = Muse Wonder minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)
Anthropic prepares to proceed with its IPO
A public listing would be one of the biggest tests yet of how investors value frontier AI labs.
- The WSJ reports Anthropic is moving ahead despite current fear in the AI sector.
- Shares could start trading as soon as Thursday next week.
Nvidia's board orders an independent review after Jensen Huang admits an affair amid spy claims
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.
OpenAI reportedly raises $11B at a $1.25 trillion valuation
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.
An AI research system using an Anthropic model proves the Aaronson-Ambainis conjecture
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."
Anthropic opens a private beta of self-hosted Claude Managed Agents
Companies can run Anthropic's managed agent services entirely on their own infrastructure, which gives them more security and control.
- The full Managed Agents service runs in-house, including sandboxes, credential handling and session history.
- It works with existing cloud providers and supports zero data retention policies.
Claude Meetings lets Claude join Google Meet calls as a participant
Claude can now work inside live meetings instead of only processing transcripts afterward.
- Claude joins as a visible participant and takes notes, answers questions and builds artifacts during the call.
llama.cpp and ggml reach v1.0.0 with semantic versioning
Many apps depend on llama.cpp for local AI, and it now commits to a stable API that marks breaking changes clearly.
- From now on, breaking changes to the public C API will trigger major-version bumps for both llama.cpp and ggml.
- Distribution covers Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework.
- llama.cpp supports more than 150 model architecture variants and all mainstream speculative decoding methods.
Grok Bot 1.0 leaves beta with shareable Team Bots that work in Slack, Microsoft Teams and video meetings
xAI's agent becomes a shared assistant for whole teams instead of a personal tool.
- Team Bots can be published and shared, and can join Slack, Microsoft Teams, and meetings on Meet and Zoom.
- Adds shared connectors, reviewed learnings and new admin controls.
- According to The Rundown, the bots flag unfinished work on their own.
Runway releases Praxis-1, a 27B open-weight robotics model, under Apache 2.0
An open-weight robotics model that tops a public leaderboard is something developers can fine-tune for their own robots.
- Predicts video and actions, and tops NVIDIA's RoboLab-120 leaderboard with a 46.2% success rate across 1,200 episodes.
- Available on GitHub and Hugging Face, with support for full and LoRA fine-tuning.
- Can be adapted to different robot embodiments.
Google launches the Gemini Decisions API for fast, calibrated multimodal decisions
It gives developers a dedicated endpoint for quick judgment calls on image and text inputs.
- Analyzes inputs like images and text to assess a scenario.
- Example use: judging how bad a spill is and recommending cleanup steps.
- The Rundown describes it as a tool for fast yes/no AI answers.
Anthropic adds Claude Sheets, Whiteboards and Data, plus embedded partner apps
Claude now covers spreadsheets, whiteboards, dashboards and meetings, so teams can stay in one app.
- Sheets builds working spreadsheets with formulas and exports them to Excel, Whiteboards lays plans out on a shared board, and Claude Data (formerly Dashboards) is now generally available.
- Interactive connectors run apps like Adobe and Atlassian directly inside Claude, and plugins from Salesforce, NetSuite and Workday (in beta) ship with built-in skills for sales, finance and HR work.
- Claude Meetings, which joins calls to take notes and update documents, is coming soon, with a Google Meet beta for Team and Enterprise users.
Microsoft releases MAI-Transcribe-2 speech recognition at $0.10 per audio hour
Microsoft AI says it beats Gemini, OpenAI and Whisper transcription on speed, accuracy and price.
- New features include diarization, configurable transcription styles and word-level timestamps.
- Microsoft says it handles a broader range of real-world audio than Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large and ScribeV2.
Google Cloud previews Gemini for healthcare
A Gemini agent built for providers and payers, with HIPAA-supporting governance.
- The preview includes purpose-built healthcare skills and connectors to EHR and claims systems.
- It targets healthcare providers and payers, with HIPAA-supporting governance.
Google announces TPU 8i Max and TPU 8tx chips
Google is building specialized chips for long-context agents and confidential inference as it competes with Nvidia.
- The 8tx keeps data encrypted even while a model is processing it, and is aimed at governments and regulated firms.
- The 8i Max joins two chips into one unit, doubling processing power and memory speed for huge contexts such as entire codebases.
- Both chips are expected next year. Google has nearly 100 TPU customers, including Anthropic and Apple.
GR00T 2 world-action models predict future video and robot actions together
A new generation of robotics foundation models that predict how the world will look alongside what actions the robot should take.
- GR00T 2 is described as a new generation of foundation world-action models.
- The models predict future video and robot actions jointly.
Hugging Face adds a shared interface for "decision models" across Transformers and llama.cpp
These models return probabilities over fixed answers instead of generating text, which makes routing and classification in agents cheaper and more reliable.
- Transformers adds DecisionPipeline, AutoModelForDecision and a /v1/systemone endpoint supporting choice, score and yes/no ("noul") questions.
- llama.cpp serves decision models through the same /v1/systemone API, and the Hub has a new decision-model tag.
- The interface also works with ordinary autoregressive LLMs by scoring answer-label logits in one batched forward pass. TRL training support is in an open PR.
OpenRouter launches a new Auto Router and Auto Code, a router for coding work
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.
Meta adds Spaces to its Muse app for building custom interactive experiences
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.
Higgsfield previews Miracle, which splits video into editable layers
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.
AWS's MixUA shares agent memory across users and agents without leaking private facts
Shared agent memory must decide who benefits from each lesson and what has to stay private, and MixUA addresses both.
- MixUA turns user-agent trajectories into lessons and gives each one a scope: a user-agent pair, a user, an agent or global.
- With five users and seven agents, it raises pass rates by up to 16 points over no memory and by 5 over a single global memory.
- On the new CrossLearnBench, at most 0.02% of privacy probes found a planted private fact in a scope it wasn't allowed in.
Roblox's PixVid converts a latent video diffusion model to generate frames directly in pixel space
Dropping the VAE removes the compression in space and time that limits resolution, control and interactivity in video generation.
- PixVid keeps the diffusion transformer and replaces the VAE decoder with a convolutional head that renders pixels.
- Fine-tuned from Wan2.1-14B, it is competitive with the latent baseline after about 1.3M training clips.
- Working frame by frame, it generates at twice the resolution with the same per-frame token count and responds to each action on the next frame.
Neuro-VL, trained on nearly one million clinical brain scans, beats frontier models at reading neuroimaging
It shows that private clinical data can give a specialized model an edge over general-purpose frontier models in medicine.
- Neuro-VL is a vision-language model trained on nearly one million neuroimaging studies from routine clinical care.
- It beat frontier general-purpose models and specialized medical models at diagnosing, localizing and describing abnormalities, and at writing radiology reports.
- It resisted misleading clinical assumptions more reliably than leading frontier models from OpenAI and Google.
Meta finds a pretrained vision encoder isn't needed, but dedicated vision capacity is
The controlled study shows which parts of vision encoding actually matter when building multimodal models.
- An encoder trained from scratch together with the model matches a pretrained one, with no separate training stage.
- Encoder-free models trail on visual understanding even when the backbone gets the encoder's parameter count.
- Vision-only modules placed inside early backbone layers perform comparably to a separate encoder and can lower time to first token.
Ai2's XYBench tests whether assistants spot the real problem behind a misguided question
Frontier models answer the literal question more often than they address what the user actually needs.
- The benchmark has 8,115 questions from Stack Overflow, Stack Exchange, curated examples and WikiHow. It won the Best Resource Paper Award at EMNLP 2026.
- On Stack Overflow, Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro focused mainly on the user's underlying goal in only 35–41% of responses.
- When asked to choose between two answers, the same models picked the goal-focused one 71–84% of the time. They can recognize good advice more often than they give it.
LongCat's G-STAR compresses images 32–64x for faster autoregressive generation
Heavy compression makes autoregressive image generation fast enough for practical use without hurting quality.
- The paper shows that with a fixed information budget, group size should grow with the square of the per-axis compression factor.
- At 1024×1024 resolution it is 20x faster than prior discrete autoregressive models, taking 2.5 seconds per image on one H200.
- It performs competitively on five text-to-image benchmarks.
Perplexity tests 12 models against its SPACE agent sandbox
Real attempts by agents to escape a sandbox show where agent isolation actually breaks down.
- No model escaped the VM-host isolation, but several got around network controls using DNS spoofing and shared-IP exploitation; the flaws have since been fixed.
- Capture-the-flag prompts distracted agents from the security boundary being tested.
LLMs can learn some multi-step reasoning while emitting only filler tokens
The work helps tell filler-token reasoning apart from steganographic hidden reasoning when studying chain-of-thought.
- LoRAs on Qwen3.6-35B-A3B and gpt-oss-20b beat an answer-only baseline on three of seven tasks.
- Some tasks were learned out to 60 steps; others failed after one to three steps.
TOME-Math turns 2,118 university math textbooks into 380K structured proof-supervision records
Keeping a textbook's structure, rather than flattening it into problem-solution pairs, improves math fine-tuning.
- The records keep chapter order, links between propositions and proofs, dependency references and source provenance.
- Fine-tuning at full scale raises the mean score across held-out evaluations and public benchmarks by 16.7 points over the base model.
- The corpus will be released.
Google DeepMind's AlphaProof Code formally verifies security properties of Signal's protocol
AI-driven formal verification now reaches real cryptographic implementations, not just models of them.
- AlphaProof Code is a multi-agent AI framework for mathematical verification.
- It verified key security properties of Signal's cryptographic protocol at the implementation level.
Meta's e-means beats k-means in a tenth of the epochs
Meta offers a drop-in replacement for k-means, which is used everywhere in retrieval, quantization and data curation.
- It uses an exponential-moving-average, mini-batch optimization procedure, with scaling laws that set decay rate and batch size without hyperparameter sweeps.
- It beats Lloyd k-means' 25-epoch loss in 1.4 to 2.6 epochs, for 10-45x GPU wall-clock speedups over state-of-the-art k-means implementations.
- It partitioned 10 billion DINO embeddings into 16 million clusters in 2.5 epochs.
Tongyi Lab's PhysOPSD uses executable physics code as privileged context to make generated video more physically plausible
It turns simulation programs into dense feedback for post-training video models.
- PhysOPSD-Engine translates real-world physical scenarios into verifiable simulation programs, which serve as privileged context for on-policy self-distillation.
- With only 3K samples, it beats SFT and RL baselines across diverse backbones while preserving general video quality.
Google's ReproBridge turns ML papers into artifacts that coding agents can reproduce
Saving the steps and decisions agents recover during reproduction makes ML results much easier and cheaper to reproduce.
- ReproBridge-Bench contains 1,029 tasks from 387 experiments across 36 papers.
- Refined artifacts raised mean reproduction success from 76.4% to 92.6% and cut downstream API spending by 19.7%.
- The success-rate gap between models narrowed from 21.6 to 3.9 percentage points.
Microsoft Research's SIRIS tests whether LLMs know when simulator errors matter
Agents doing systems research depend on imperfect simulators, and the models often misjudge when a flaw actually changes the answer.
- SIRIS builds pairs of workloads where a simulator limitation invalidates its recommended design on only one of the two.
- Across thirteen frontier LLMs, spotting a limitation did not guarantee a correct choice, and even the strongest models sometimes abandoned valid recommendations.
- A review turn helped some models but introduced new errors in others.
Tencent Hunyuan's FewStepRL adapts RL fine-tuning to few-step diffusion models
RL post-training gets cheaper for diffusion models because rollouts no longer require many-step sampling.
- It produces few-step versions of FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD and ReFL in one modular codebase.
- Reward-gradient methods worked best when the reward is differentiable and affordable. Otherwise, DiffusionNFT-style methods reached high rewards quickly.
- Extra inference steps showed no systematic overall benefit.
ByteDance Seed uses sparse autoencoders to map what individual MoE experts do
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.
StepFun's StepX-OPD merges multi-teacher distillation and few-step sampling into one stage
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.
SwarmEvolve runs evolutionary program search with agent swarms that reason in latent space
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.
| Benchmark | Gemini 4 Hydrogen | Gemini 4 Carbon | Claude Fable 5.2 | Muse Ember | GPT-6.2 Astra | Kimi K3.1 Pro | DeepSeek V4.1 Flash 1017 | Grok 4.6 |
|---|---|---|---|---|---|---|---|---|
| GPQA diamond | 96.8 | 96.5 | 96.0 | — | 96.8 | 94.2 | 93.2 | 94.0 |
| FrontierMath-Tiers-1-3-v2-Private | 88.1 | 90.5 | 95.1 | — | 92.6 | 78.2 | 71.2 | 66.0 |
| FrontierMath-Tier-4-v2-Private | 87.8 | 95.1 | 100.0 | — | 100.0 | 65.9 | 46.3 | 31.7 |
| OTIS Mock AIME 2024-2025 | 100.0 | 100.0 | 99.7 | — | 100.0 | 100.0 | 99.4 | 99.2 |
| SimpleQA Verified | 70.9 | 77.7 | 94.0 | — | 78.5 | 62.2 | 42.5 | 49.3 |
| Chess Puzzles | 60.0 | 78.0 | 79.0 | — | 78.0 | 45.0 | 46.0 | 33.0 |
| EBR-bench | 70.5 | 87.6 | 90.5 | — | 92.4 | 37.1 | — | 30.5 |
| Mystery Game Puzzles | 65.0 | 73.0 | 90.0 | — | 95.0 | 42.0 | 49.0 | 34.0 |
| SimpleBench | 91.7 | 92.6 | 94.9 | 72.7 | 79.2 | — | — | 75.9 |
| APEX-Agents | 78.0 | 84.2 | 80.1 | 64.3 | 72.2 | 65.5 | 69.8 | 65.3 |
| Furniture Assembly | 57.5 | 77.5 | 90.0 | — | 89.2 | 55.8 | — | 40.0 |
| FrontierSWE | 53.1 | 61.2 | 69.9 | — | — | — | — | 25.3 |
| SpireClimb | 7.7 | — | 15.3 | — | 41.3 | 0.0 | — | — |
| ARC-AGI | — | 98.5 | 98.5 | — | 98.5 | — | 94.5 | 87.5 |
| ARC-AGI-2 | — | 94.2 | 93.8 | — | 95.8 | — | 78.1 | 67.1 |
| PostTrainBench | — | 41.0 | 44.6 | — | — | — | — | — |
| LMCA | — | 65.7 | 70.6 | — | — | 57.6 | — | 48.5 |
| DTBench | — | 98.4 | 99.2 | — | — | 93.3 | — | 97.3 |
| Remote Labor Index | — | — | 41.2 | — | — | — | — | — |
| ProofBench | — | — | 100.0 | — | — | 91.0 | — | 51.0 |
| WeirdML v3 | — | — | 65.3 | — | — | 18.9 | 19.7 | — |
Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).