AI Digest

Category

METR estimates Anthropic's AI progress has roughly doubled, mostly driven by its own AI

The Rundown AI

An outside evaluator with 11 weeks of inside access says the AI-building-AI loop is already running at a frontier lab.

  • METR's report, based on 11 weeks of access to Anthropic's internal data and staff interviews, estimates a 1.3x to 3x speedup over the company's 2024-25 pace, crediting AI for about 80% of the faster progress.
  • Anthropic redacted most of the strongest evidence; safety lead Sam McCandlish said this was to avoid "handing out a roadmap for others to go faster."
  • The report revealed two internal models ahead of Claude Opus 5.5, including a "Model 4" that cleared 50% on all 15 tested benchmarks, and METR found no clear bottlenecks.
Read on metr.org →

OpenAI releases GPT-6 Cyber, a model for finding and exploiting software flaws

The Rundown AI

A frontier lab is now shipping a dedicated offensive-security model that outperforms its general flagship on cyber tasks.

  • GPT-6 Cyber is built for hunting and exploiting software vulnerabilities and beats OpenAI's main GPT-6 Astra model on most cyber tasks.
  • In testing, the model found 81 serious new bugs in Firefox's browser.
Read on openai.com →
Benchmarks
BenchmarkGPT-6 AstraClaude Fable 5.2Gemini 4 CarbonMuse WonderKimi K3.1 ProΔ vs best other
GPQA diamond95.896.096.5—94.2-0.7
FrontierMath-Tiers-1-3-v2-Private93.795.190.582.178.2-1.4
FrontierMath-Tier-4-v2-Private97.6100.095.170.765.9-2.4
OTIS Mock AIME 2024-2025100.099.7100.099.4100.0+0.0
SimpleQA Verified75.694.077.7—62.2-18.4
Chess Puzzles76.079.078.0—45.0-3.0
EBR-bench76.290.587.631.437.1-14.3
Mystery Game Puzzles84.090.073.051.042.0-6.0
MirrorCode60.0——45.643.3+14.4
ARC-AGI98.598.598.5——+0.0
Balrog68.3—————
GSO-Bench79.4—————
SimpleBench86.594.992.6——-8.4
WeirdML93.6—————
APEX-Agents64.780.184.266.165.5-19.5
ARC-AGI-295.093.894.2——+0.8
PostTrainBench41.944.641.024.2—-2.7
Remote Labor Index20.841.2———-20.4
ProofBench99.0100.0—84.091.0-1.0
FrontierCode53.3—————
LMCA64.470.665.756.857.6-6.2
DTBench97.399.298.495.293.3-1.9
Furniture Assembly80.090.077.554.255.8-10.0
FrontierSWE65.569.961.2——-4.4
WeirdML v354.065.3—24.918.9-11.3
SpireClimb34.015.3——0.0+18.7

Δ = GPT-6 Astra minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)

Thinking Machines releases Inkling 2 Small, a 594B MoE with Apache 2.0 weights coming

TLDR AI

A large mixture-of-experts model with vision and agentic building skills is heading to open weights after safety testing.

  • Inkling 2 Small is a Mixture-of-Experts model with 594B total parameters and 17B active, supporting image inputs and variable effort.
  • Its vision and agentic capabilities let it use visual feedback to iteratively build and refine websites, games and 3D designs with interactive elements.
  • Initially available through Tinker with closed weights; full weights arrive under Apache 2.0 next week after external safety testing and red-teaming.
Read on thinkingmachines.ai →
Benchmarks
BenchmarkInklingGPT-6 AstraClaude Opus 5.5Gemini 4 ArgonMuse WonderΔ vs best other
GPQA diamond88.395.890.696.4—-8.1
FrontierMath-Tiers-1-3-v2-Private33.393.791.286.082.1-60.4
FrontierMath-Tier-4-v2-Private4.997.695.078.070.7-92.7
OTIS Mock AIME 2024-202588.9100.0100.0100.099.4-11.1
SimpleQA Verified40.375.672.276.0—-35.7
Chess Puzzles21.076.067.059.0—-55.0
ARC-AGI79.598.598.598.5—-19.0
SimpleBench50.0—————
WeirdML32.393.6———-61.3
APEX-Agents33.864.773.582.466.1-48.6
ARC-AGI-236.595.093.391.2—-58.5
ProofBench0.099.0100.099.084.0-100.0
FrontierCode14.053.354.647.5—-40.6
LMCA37.664.468.263.756.8-30.6
DTBench87.597.398.997.695.2-11.4
FrontierSWE4.165.562.355.0—-61.4
WeirdML v34.454.049.841.125.2-49.6

Δ = Inkling minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)

Anthropic says it may be near its internal definition of recursive self-improvement

TLDR AI

An IPO filing disclosure puts a frontier lab's own self-improvement threshold on the record alongside its financials.

  • Anthropic disclosed in its updated IPO filing that it may soon reach its internal definition of recursive self-improvement, where AI systems help develop increasingly capable models.
  • The amended prospectus reports a 1,200% revenue increase to $33.24 billion for the first nine months of 2026, with a net loss of $104 billion.
  • Compute and infrastructure commitments total approximately $895.4 billion.
Read on techcrunch.com →

Ripple-65B Preview tops open-weight models on agentic coding for $32M in training

TLDR AI

A relatively cheap open model beats far larger US open-weight releases on TerminalBench 2.1.

  • Ripple-65B Preview reaches state-of-the-art performance on TerminalBench 2.1, surpassing larger open models including Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B and Nvidia Nemotron 3 Ultra 550B.
  • It was trained for $32 million, one-tenth the estimated cost of the previous open-weight SOTA on agentic coding.
  • The model is available on Hugging Face, Together AI, GMI Cloud and OpenRouter.
Read on ricursive.com →
Benchmarks
BenchmarkMuse Spark OpenClaude Fable 5.2GPT-6.2 AstraGemini 4 CarbonKimi K3.1 ProΔ vs best other
GPQA diamond88.696.096.896.594.2-8.2
FrontierMath-Tiers-1-3-v2-Private69.895.192.690.578.2-25.3
FrontierMath-Tier-4-v2-Private31.7100.0100.095.165.9-68.3
OTIS Mock AIME 2024-202598.199.7100.0100.0100.0-1.9
SimpleQA Verified56.694.078.577.762.2-37.4
ARC-AGI89.098.598.598.5—-9.5
APEX-Agents33.480.172.284.265.5-50.8
ARC-AGI-257.593.895.894.2—-38.3

Δ = Muse Spark Open minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)

DeepSeek plans to pretrain its next model on Huawei Ascend chips

TLDR AI

The run would be the first real test of whether Chinese accelerators can fully pretrain a frontier LLM.

  • DeepSeek plans to use Huawei's Ascend AI chips to pretrain DeepSeek V5; Ascend chips have previously been used mainly for inference.
  • A successful run could reduce Chinese developers' dependence on Nvidia hardware and strengthen Huawei as an alternative for advanced AI computing.
Read on geopolitechs.org →

Marvell agrees to buy Lumentum for $115B in all-stock deal

The Rundown AI

One of tech's largest takeovers ever targets the optical components underpinning AI data centers.

  • Marvell agreed to a $115B all-stock deal to acquire Lumentum, which makes lasers for AI data centers.
  • The Rundown describes it as one of the largest takeovers in tech history.
Read on investor.marvell.com →

Meta acquires AI coding startup Conductor

TLDR AI

Meta picks up tooling for running many coding agents at once as agent orchestration becomes a battleground.

  • Conductor lets developers run multiple AI coding agents in parallel workspaces.
  • The startup's platform helps developers manage multiple AI agents at the same time, per The Rundown.
Read on conductor.build →

OpenAI to acquire Pierre Computer Company

TLDR AI

The deal folds the startup behind Codex's code review tools directly into OpenAI's developer stack.

  • OpenAI will acquire Pierre Computer Company to enhance Codex's developer tools, including Code Storage and other open-source projects.
  • Pierre built the code review tools inside Codex, and OpenAI plans to add its Code Storage service to the platform.
Read on openai.com →

Apple opens Siri AI to all developers in a preview

The Rundown AI

A rare opening of Apple's assistant, though App Store submissions don't start until spring.

  • Apple released a developer preview of Siri AI in its latest betas, letting apps build their own actions and teach Siri multi-step jobs through new "skills."
  • App versions stay with a developer's internal testers for now; App Store submissions start in spring and developer rules will come later.
  • Siri AI is Apple's rebuilt assistant, able to see a user's screen and read messages, and has been a waitlisted beta since iOS 27 shipped in September.
Read on developer.apple.com →

Data center ballot measures lose almost everywhere as Democrats take the House

The Rundown AI

The US midterms turned into a referendum on the AI buildout, with direct consequences for data center expansion and AI oversight.

  • Data center opponents have taken 38 of 39 fully counted local measures, with only Harwood, N.D., voting the other way.
  • In NBC's exit poll, 76% of voters would oppose a data center in their own community and 57% expect AI to hurt society more than help.
  • Goldman Sachs reportedly sees 45% of planned US data center capacity at risk of temporary freezes.
Read on datacentersontheballot.com →

Google ties school Search access to its AI features

The Rundown AI

Districts that want AI out of classrooms may have to give up Google Search for students too.

  • K-12 student accounts will start with Search, Gemini and Gemini Notebook all off early next year, with students at AI-free schools seeing a notice that it's off.
  • Education VP Chris Phillips told The New York Times that "Search and A.I. are deeply integrated."
  • Ex-FTC Chair Lina Khan said Google is using "its search monopoly to coerce schools"; Google defended the move on X as "more choice, not less."
Read on blog.google →

OpenAI ships Images 2.6 with GIF generation

The Rundown AI

OpenAI's image model family adds motion for the first time while getting faster.

  • Images 2.6 rolls out across ChatGPT and Codex, and can animate a typed idea, photo, or sketch into a GIF, or create animations from scratch.
  • Two API versions are available to developers: Sunburst for higher-quality work and Flare for lighter, faster jobs.
  • OpenAI says the model generates images 50% faster than previous versions, and 2.6 edges past its predecessor for the top spot on Design Arena.
Read on openai.com →

Google rolls out Gemini Omni 1.2 Flash with a 30% price cut

The Rundown AI

Google's video model gets cheaper and generates longer clips.

  • Pricing drops 30% to $0.07 per second at 720p.
  • Maximum clip length extends from 10 seconds to 15 seconds.
Read on blog.google →

Anthropic adds a Personal Finance tool to Claude

The Rundown AI

Claude can now connect directly to a user's financial accounts.

  • The new tool lets Claude connect to a user's financial accounts, answer questions about them, and provide tailored information.
Read on claude.com →

Perplexity launches Teammates, project-owning agents on Perplexity Computer

The Rundown AI

Perplexity is betting that each worker will supervise many persistent agents rather than one assistant.

  • Teammates are AI agents on the Perplexity Computer platform that each own an ongoing project.
  • CEO Aravind Srinivas said he anticipates "every org having hundreds of teammates working for each person."
Read on perplexity.ai →

Tencent launches Hy4, an open-source flagship with visual understanding

TLDR AI

Another major Chinese open-weights model, this one aimed at document, chart and computer-use work.

  • Hy4 is a large language model with visual understanding; Tencent plans to release weights and system cards in the coming days.
  • It supports fine-grained image processing, complex chart analysis and mathematical visual reasoning.
  • The model can interact with both command-line interfaces and GUIs to perform work and everyday tasks.
Read on tencent.com →
Benchmarks
BenchmarkHy4DeepSeek-V4.1-Flash-1017Kimi K3.1 ProGLM-5.3Hy4 PreviewKimi K3Δ vs best other
Financial analysis blind evaluation (213 financial professionals, 579 real-world tasks)3.03.02.72.7——+0.1
WorkBuddy (average score)83.2——————
WorkBuddy (pass rate %)96.2——————
Hallucination rate (%, lower is better)7.2———12.9—-5.7
Unsupported-assumption error rate (%, lower is better)5.6———23.1—-17.5
GPQA Diamond————92.3——
SWE-bench Pro————65.7——
Deep SWE————64.3——
Engineering blind evaluation (163 experts, 203 engineering tasks)———2.93.02.9—

Δ = Hy4 minus the best other model listed. Sources: huggingface.co · www.tencent.com

Moonshot AI releases Kimi K3.1 Pro

TLDR AI

A 3T-class open-weight multimodal model with a million-token context window.

  • Kimi K3.1 Pro is the latest in the 3T-class models and excels in multimodal tasks.
  • It improves outputs in coding, 3D creation and factual accuracy, with a million-token context window.
Read on huggingface.co →
Benchmarks
BenchmarkKimi K3.1 ProClaude Fable 5.2GPT-6.2 AstraGemini 4 CarbonMuse WonderΔ vs best other
GPQA diamond94.296.096.896.5—-2.6
FrontierMath-Tiers-1-3-v2-Private78.295.192.690.582.1-16.9
FrontierMath-Tier-4-v2-Private65.9100.0100.095.170.7-34.1
OTIS Mock AIME 2024-2025100.099.7100.0100.099.4+0.0
SimpleQA Verified62.294.078.577.7—-31.8
Chess Puzzles45.079.078.078.0—-34.0
EBR-bench37.190.592.487.631.4-55.3
Mystery Game Puzzles42.090.095.073.051.0-53.0
MirrorCode43.3—67.8—45.6-24.5
APEX-Agents65.580.172.284.266.1-18.7
ProofBench91.0100.0——84.0-9.0
LMCA57.670.6—65.756.8-13.0
DTBench93.399.2—98.495.2-5.9
Furniture Assembly55.890.089.277.554.2-34.2
WeirdML v318.965.3——24.9-46.4
SpireClimb0.015.341.3——-41.3

Δ = Kimi K3.1 Pro minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)

NVIDIA adds four open-weight Cosmos 3 models for interactive world simulation

Hugging Face Blog

Open checkpoints for driving and robot world models, with post-training recipes for your own sensor rig.

  • The four post-trained Cosmos 3 Nano checkpoints — Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual — each have 16B trainable parameters under the OpenMDW1.1 license.
  • Transfer-Auto generates synchronized video across up to 11 camera views, including fisheyes, and supports joint RGB–LiDAR generation; Sim-Auto is its causally distilled streaming variant.
  • Sim-Bimanual simulates how an environment evolves from a two-arm robot's actions for YAM and AgiBot embodiments.
Read on huggingface.co →

Google releases an AlphaEvolve implementation and its discoveries on GitHub

TLDR AI

The Gemini-powered algorithm-design agent behind data center and chip efficiency gains is now public code.

  • AlphaEvolve pairs Gemini models with automated evaluators that verify answers, as a coding agent for designing advanced algorithms.
  • It has been used to improve the efficiency of Google's data centers, chip design and AI training process.
  • The repo contains a minimal implementation plus code and data related to the solutions and discoveries it contributed to.
Read on github.com →

Grok Bot starts making phone calls for paid users

The Rundown AI

An agent that sits on hold and negotiates bills, with users made the legally responsible "caller of record."

  • Grok Bot can now sit on hold, book appointments and negotiate bills on U.S. numbers, with early examples including a Comcast bill negotiation and a pizza order.
  • If a rep asks something the bot can't answer, users get a mid-call message and type a reply for it to pass along while following a live transcript.
  • Terms make users the "caller of record," "solely responsible" for calls and phone-law compliance, with bots disclosing they are a "virtual assistant."
Read on x.com →

Anthropic ships Claude Message Threads in public beta

TLDR AI

Agents can stop resending entire conversation histories, cutting latency on very long contexts.

  • Message threads let agents send only new messages instead of resending their entire conversation history.
  • The feature reduced median time to first token by roughly 35% for requests with 500K–1M tokens, without changing token usage.
Read on claude.com →

Google opens a Gemini Notebook API in early access

TLDR AI

A REST interface turns web pages and videos into programmatically managed notebooks.

  • The API lets developers create and manage notebooks and sources, turning web pages, YouTube videos, plain text and Markdown into persistent notebooks.
  • The API and an MCP server are in an early access Public Preview.
  • Only personal Google Accounts are supported during the preview; Workspace accounts are not yet supported.
Read on developers.google.com →

Gradium's Phonon brings high-quality TTS fully on-device

Hugging Face Blog

A 65–75 MB speech model tops the Open TTS Leaderboard on speed and leads intelligibility in two languages.

  • Phonon comes in two sizes, about 65 MB and 75 MB, speaking English, French, German, Spanish and Portuguese with 100+ built-in voices; Phonon 31m is Apache 2.0 and Phonon 40m is CC BY-NC.
  • It ranks first on intelligibility in German and Spanish and in the top three in every language evaluated, and tops the CPU streaming benchmark on time-to-first audio.
  • It runs via a new ptts framework with 10+ CPU, GPU and NPU backends, and in Transformers from v5.21; it was downloaded over 200,000 times in its first two weeks.
Read on huggingface.co →

CrowdStrike adds an "AI Kill Switch" for rogue agents

The Rundown AI

An emergency stop sitting outside the model, as enterprises worry about agents holding company credentials.

  • The switch targets an agent's link to its AI model, not the model itself, while Falcon Guardian separately monitors tools like Claude Code and Codex.
  • A human admin must pull the trigger and every use is logged; shutting agents down directly on work devices and in the cloud is "coming soon."
  • CrowdStrike is also buying monitoring firm Fiddler AI and security tester Dreadnode for a reported $495M, and added Claude Cowork protection in beta.
Read on crowdstrike.com →

Pangram Image 1.0 localizes AI-generated regions in images and video

TLDR AI

AI detection that outputs a heatmap rather than a single verdict, and holds up through edits and compression.

  • Pangram Image 1.0 produces a heatmap localizing exactly where AI appears in images and is robust to edits, compression and low resolutions.
  • A Chrome extension is available alongside API support for images and video.
Read on pangram.com →

Ai2 opens ToyBlocks, a controlled simulation world for robot learning and evaluation

Hugging Face Blog

A deliberately narrow block-manipulation world lets researchers trace why robot policies fail instead of only scoring success rates.

  • ToyBlocks auto-generates tasks, skills and demonstrations without teleoperation, running hundreds of plans in parallel on a single GPU at up to 24x real-time; the initial release includes 85,336 long-horizon demonstrations, about 1,600 hours of execution data and 144 million frames.
  • Generalization tests found that ManiFlow, π0.5 and MolmoAct 2 generalize poorly to new spatial configurations despite large-scale pretraining; a hierarchical Molmo 2 + ManiFlow system scored 55.5% in simulation and 50.0% zero-shot on a real Franka FR3 arm.
  • In agentic tests, a system using Claude Opus 5.5 over MCP tools solved 30/30 classical block-stacking tasks but reached only 63% on search-and-spell, with errors from look-alike letters (13%) and wrong letter orientation (15%); agent-written skill descriptions raised success from 24% to 63%, beating a human expert's 36%.
Read on huggingface.co →

MIT shows GPT-6 Astra improving another model by editing its weights directly

The Rundown AI

A concrete, if narrow, demonstration of a model upgrading another model without any training run.

  • Alibaba's small Qwen model scored 7.9% in math until Astra changed 0.016% of its weights to cut a repetition habit, reaching 48.1%.
  • The same approach barely moved math scores for Google's smaller Gemma 3, which lacked a single flaw like Qwen's repetition.
  • In smaller tests Astra named the game a model had learned from its weights alone and wrote a simple Breakout player from scratch; TLDR also flags commentary arguing the "read and write weights" framing is misleading and not superhuman.
Read on rikiyat.github.io →

OpenAI reports validating and patching 948 Linux privilege-escalation bugs in one run

TLDR AI

The first large-scale result from OpenAI's autonomous vulnerability pipeline behind GPT-6 Cyber.

  • OpenAI developed an autonomous pipeline for discovering, exploiting and patching security vulnerabilities.
  • In a single run, its cybersecurity model validated and generated patches for 948 Linux local privilege escalation vulnerabilities.
Read on openai.com →

OpenAI publishes LifeVal, a benchmark for everyday errands

TLDR AI

A year-old internal eval for the chores people actually ask agents to do, now public.

  • LifeVal grades models on tasks like booking appointments, buying groceries and filling out government forms, with task areas drawn from a poll of 4,000 people and global time-use surveys.
  • OpenAI's models went from an average score of 49.3% (GPT-5.2) to 80.6% (GPT-6 Astra), though only 0.6 points of that gain came after July's GPT-5.6.
  • Astra's best recorded scores hit 100% on 10 of 16 listed tasks, with LSAT help lowest at 88.7%.
Read on openai.com →

Progress on cryptographic proofs of which model produced an output

TLDR AI

zkSNARK-style proofs of inference could underpin compliance checks on what model actually answered.

  • Verifying which model produced a specific output is almost impossible today; cryptographic proofs of inference would let providers prove an output came from a specific model.
  • The write-up details progress on zkSNARKs, a type of proof the authors argue is uniquely fit for provable inference.
  • Separately, Anthropic and university researchers used Claude to help build FlashProve, an open-source way to prove which model wrote an answer via cryptographic proofs.
Read on anthropic.com →

NVIDIA's VFlash decodes visual plans in one pass instead of token by token

Hugging Face Daily Papers

Predicting whole visual primitives jointly makes robot trajectory planning far cheaper and more accurate.

  • A lightweight drafter generates an entire visual primitive — points, boxes, variable-length trajectories — in one forward pass, adapting to its type, dimensionality and length.
  • VFlash trains on annotated primitives and accepts structurally valid predictions without requiring agreement with an autoregressive model's exact tokens.
  • Preliminary single-rollout measurements indicate ~25x faster 3D point motion forecasting, and a world action model conditioned on VFlash plans hits 70.2% average success across four MolmoSpaces suites versus 62.6% for the baseline.
Read on huggingface.co →

Google's Muir reallocates linear-attention gating for long-context recall

Hugging Face Daily Papers

A sensitivity analysis of gated delta-rule models yields cheaper gates and much longer extrapolation.

  • The authors show state updates are far less sensitive to erasure than to writes, so the erase gate can be parameterized at lower rank, with savings reinvested in a wider feed-forward network.
  • Across 166M–1.34B parameters, Muir lowers perplexity by 1.8–2.0% over GDN-2 and 2.5–3.6% over Transformer, and raises commonsense accuracy by up to 1.9 points.
  • At 1.34B it extrapolates 2,560x from a 4K training length to 10M-token streams without perplexity drift and keeps near-100% passkey retrieval to 1M tokens, while matching GDN-2 throughput.
Read on huggingface.co →

Microsoft Research's Hybrid Computer Use blends code, accessibility and GUI channels

Hugging Face Daily Papers

Letting agents pick among programmatic, semantic and visual interfaces beats screenshot-only computer use.

  • HCU unifies programmatic interaction via code and CLI tools, semantic interaction via accessibility trees, and visual GUI interaction, with guards against stale-state actions when switching channels.
  • Its accessibility tools return compact observations, batch actions and report effects as AX diffs, cutting round trips.
  • On OSWorld 2.0, HCU achieves the highest Partial and Binary scores among compared systems, and all evaluated models improve with it though they allocate work across channels differently.
Read on huggingface.co →

WeLM-Image unifies generation, editing and multi-reference synthesis on one GPU

Hugging Face Daily Papers

A sparse 31.7B image foundation model built for single-GPU deployment inside WeChat-scale apps.

  • A Mixture-of-Transformers backbone with per-path MoE modules stores 31.7B parameters and activates about 5.4B generation parameters per target latent token.
  • The 1536-resolution model ranks first among evaluated open-source models on PRISM-Bench, Qwen-Image-Bench and UnicBench.
  • MeanFlow few-step distillation yields WeLM-Image-Turbo, generating a 1024×1024 image in 4.98s on a single NVIDIA H20 using eight guidance-free steps.
Read on huggingface.co →

Oxford argues for "cognitive symbionts" alongside autonomous agents

Hugging Face Daily Papers

A position paper reframes Physical AI around augmenting a specific person rather than building an independent machine.

  • The paper proposes AI-augmented human intelligence (AHI), asking how much intelligence a machine can enable in a particular person, next to AGI's question of how intelligent a machine can be.
  • Its conceptual core is the "cognitive gap" between what a person could bring to a situation and what they can hold in working memory, with a design space organized as a symbiosis loop enclosed by a personal runtime and trust.
  • The authors treat governance as part of the paradigm, citing risks from covert assistance in examinations to pressure on everyone to adopt it.
Read on huggingface.co →

Open-HumanScale scales egocentric human video pretraining to 50,000 hours

Hugging Face Daily Papers

An open corpus and recipe yield a log-linear scaling law from human video hours to robot success rates.

  • Open-HumanScale-50K is a curated egocentric video corpus densely annotated with hand motion and natural language, supporting an empirical log-linear scaling law between pretraining hours and downstream robot success.
  • HumanWAM-Vanilla, pretrained only on human video, reaches 92.08% success on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random after robot post-training.
  • Joint human–robot pretraining with HumanWAM-Goal reaches 92.70% Clean2Clean and 84.60% Clean2Random, and shows emergent in-context adaptation to unseen visual goals without parameter updates.
Read on huggingface.co →
BenchmarkGPT-6 AstraInkling 2 SmallMuse Spark OpenKimi K3.1 ProClaude Fable 5.2Gemini 4 CarbonDeepSeek V4.1 Flash 1017Grok 4.6
GPQA diamond95.8—88.694.296.096.593.294.0
FrontierMath-Tiers-1-3-v2-Private93.7—69.878.295.190.571.266.0
FrontierMath-Tier-4-v2-Private97.6—31.765.9100.095.146.331.7
OTIS Mock AIME 2024-2025100.0—98.1100.099.7100.099.499.2
SimpleQA Verified75.6—56.662.294.077.742.549.3
Chess Puzzles76.0——45.079.078.046.033.0
EBR-bench76.2——37.190.587.6—30.5
Mystery Game Puzzles84.0——42.090.073.049.034.0
MirrorCode60.0——43.3———41.1
ARC-AGI98.596.589.0—98.598.594.587.5
Balrog68.3———————
GSO-Bench79.4———————
SimpleBench86.5———94.992.6—75.9
WeirdML93.6——————67.3
APEX-Agents64.746.833.465.580.184.269.865.3
ARC-AGI-295.081.757.5—93.894.278.167.1
PostTrainBench41.9———44.641.0——
Remote Labor Index20.8———41.2———
ProofBench99.0——91.0100.0——51.0
FrontierCode53.3——————48.0
LMCA64.4——57.670.665.7—48.5
DTBench97.3——93.399.298.4—97.3
Furniture Assembly80.0——55.890.077.5—40.0
FrontierSWE65.525.1——69.961.2—25.3
WeirdML v354.0——18.965.3—19.7—
SpireClimb34.0——0.015.3———

Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).