#1
Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It
A small LoRA is enough to stop transformers from quitting a chain of thought early.
Top 5 Signalcurated
The curator's five highest-signal stories, merged across sources.
#2
Introducing GPT-6.1 Sol
GPT-6.1 Sol prices near-Astra coding and computer-use at one-fifth of Astra's API token rates.
#3
Gemini 4 Argon: our next era of frontier intelligence
Gemini 4 Argon is DeepMind's new flagship, with a 1 million token output and first access for cyber defenders.
#4
Transformers now runs llama.cpp quants
Transformers can now load llama.cpp quantization formats directly.
Benchmarksrankings
Cross-benchmark aggregation across public leaderboards. Sources that stop updating are auto-excluded — see below.
31 models ranked across 11 benchmarks — aggregated from multiple sources. Top composite: Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5.5.
Data as of 2026-10-19 · live sources: Artificial Analysis, Vellum.
Composite Ranking
Ranked by head-to-head record: each pair of models is compared only on benchmarks BOTH were tested on (near-ties count as draws), so a model gains nothing from appearing on more or easier leaderboards. Avg percentile is shown for context. Only models with 3+ benchmarks are ranked; the rest appear in the heatmap below.
| # | Model | H2H (W–L–T) | Avg Pctl | # Bench | Sources |
|---|---|---|---|---|---|
| T1 | Claude Opus 5.5· | 26–0–0 | 94th | 6 | Artificial Analysis, Vellum |
| T1 | Claude Fable 5.1· | 25–1–0 | 87th | 6 | Artificial Analysis, Vellum |
| 3 | Claude Sonnet 5.5· | 24–2–0 | 81st | 5 | Artificial Analysis, Vellum |
| T4 | Muse Wonder Full· | 22–3–1 | 80th | 5 | Artificial Analysis |
| T4 | Gemini 4 Argon· | 22–3–1 | 71st | 5 | Artificial Analysis |
| 6 | GPT-6 Astra· | 21–5–0 | 74th | 7 | Artificial Analysis, Vellum |
| T7 | GPT-6.1 Sol· | 19–6–1 | 68th | 6 | Artificial Analysis |
| T7 | Kimi K3· | 18–6–2 | 66th | 7 | Artificial Analysis, Vellum |
| T7 | Muse Wonder· | 18–7–1 | 64th | 5 | Artificial Analysis |
| 10 | MiMo-V2.6-Pro▲1 | 17–9–0 | 67th | 5 | Artificial Analysis |
| T11 | Step 5 Preview▲1 | 15–10–1 | 55th | 6 | Artificial Analysis |
| T11 | Gemini 3.8 Flash▲1 | 14–10–2 | 60th | 7 | Artificial Analysis |
| T13 | Grok 4.7▲1 | 13–11–2 | 49th | 5 | Artificial Analysis |
| T13 | GLM-5.3▲1 | 12–12–2 | 44th | 6 | Artificial Analysis |
| T15 | Claude Haiku 5.5▲1 | 11–14–1 | 47th | 5 | Artificial Analysis |
| T15 | Qwen3.8 Max▲1 | 10–15–1 | 47th | 7 | Artificial Analysis |
| T17 | DeepSeek V4.1 Flash 1017new | 9–15–2 | 47th | 6 | Artificial Analysis |
| T17 | Muse Spark Opennew | 8–14–4 | 40th | 7 | Artificial Analysis |
| 19 | GLM-5.3-Flash▼1 | 8–17–1 | 33rd | 6 | Artificial Analysis |
| 20 | GPT-6 Luna▼1 | 7–19–0 | 37th | 6 | Artificial Analysis |
| T21 | MiniMax-M3▼1 | 5–20–1 | 38th | 10 | Artificial Analysis |
| T21 | Mistral Large 4 Preview· | 5–21–0 | 25th | 6 | Artificial Analysis |
| T23 | Beam 1.0new | 3–21–2 | 17th | 5 | Artificial Analysis |
| T23 | Qwen3.8 27B▼1 | 3–22–1 | 20th | 7 | Artificial Analysis |
| T25 | K2 Horizon 375B A23B▼1 | 1–24–1 | 13th | 6 | Artificial Analysis |
| T25 | Inkling▼1 | 1–24–1 | 10th | 7 | Artificial Analysis |
| 27 | Nemotron 3.5 Super VLnew | 0–26–0 | 7th | 5 | Artificial Analysis |
Benchmark Heatmap
Darker = higher score within each benchmark column. Columns use different scales, so compare within a column, not across.
| Model | AA Intelligence Index | Humanity's Last Exam | SciCode | AA Long Context Reasoni… | AA Omniscience | MMMU-Pro | GPQA Diamond | IFBench | SWE-Bench | Tau2 | Terminal-Bench Hard |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 58 | 68 | 67 | 85 | 46 | 88 | — | — | — | — | — |
| Claude Fable 5.1 | 53 | 65 | 63 | 85 | 44 | — | 94 | — | — | — | — |
| Claude Sonnet 5.5 | 56 | 64 | 61 | 83 | 32 | — | — | — | — | — | — |
| Muse Wonder Full | 56 | 54 | 63 | 84 | 30 | — | — | — | — | — | — |
| Gemini 4 Argon | 53 | 57 | 62 | 80 | 42 | — | — | — | — | — | — |
| GPT-6 Astra | 53 | 57 | 56 | 81 | 43 | 87 | 96 | — | — | — | — |
| GPT-6.1 Sol | 52 | 53 | 54 | 83 | 42 | 86 | — | — | — | — | — |
| Kimi K3 | 44 | 56 | 60 | 89 | 20 | 80 | 94 | — | — | — | — |
| Muse Wonder | 52 | 50 | 60 | 81 | 26 | — | — | — | — | — | — |
| MiMo-V2.6-Pro | 46 | 49 | 61 | 86 | 8 | — | — | — | — | — | — |
| Step 5 Preview | 44 | 46 | 59 | 88 | 16 | 76 | — | — | — | — | — |
| Gemini 3.8 Flash | 41 | 48 | 57 | 81 | 30 | 86 | 95 | — | — | — | — |
| Grok 4.7 | 46 | 43 | 57 | 77 | 32 | — | — | — | — | — | — |
| GLM-5.3 | 45 | 42 | 59 | 80 | 14 | — | 92 | — | — | — | — |
| Claude Haiku 5.5 | 43 | 44 | 55 | 83 | 11 | — | — | — | — | — | — |
| Qwen3.8 Max | 45 | 43 | 52 | 80 | 12 | 83 | 93 | — | — | — | — |
| DeepSeek V4.1 Flash 1017 | 45 | 43 | 56 | 83 | 5 | 81 | — | — | — | — | — |
| Muse Spark Open | 39 | 44 | 58 | 78 | 25 | 81 | 89 | — | — | — | — |
| GLM-5.3-Flash | 42 | 40 | 52 | 80 | 8 | — | 91 | — | — | — | — |
| GPT-6 Luna | 38 | 38 | 55 | 83 | 1 | 80 | — | — | — | — | — |
lowerhigher · scaled within each column; bold = column leader
Top by Downloadstools
HuggingFace trending models, ranked by recent downloads.
#1 · Qwen/Qwen3.8-27B
6,821,761 downloads · 16,906 likes · $1.49/M out
#2 · Lightricks/LTX-2.5
1,626,951 downloads · 6,202 likes
#3 · abenzerps/Qwen-Image-2.1-Uncensored-GGUF
1,553,744 downloads · 3,005 likes
#4 · Qwen/Qwen-Image-2.1
90,003 downloads · 2,920 likes
#5 · Cloudflare/clef-flash
6,372 downloads · 394 likes
#6 · Cloudflare/clef
4,214 downloads · 1,084 likes
#7 · convaiinnovations/laya
3,752 downloads · 5,115 likes
#8 · Contrastive-LM/CLM-v0.1-8B
3,445 downloads · 698 likes
#9 · PSRben/VisionHOPE
1,516 downloads · 393 likes
#10 · Aleph-Alpha/Kolibri-1
1,135 downloads · 323 likes
GitHub Trendingtools
Trending repos on GitHub today, ranked by stars.
#1 · DietrichGebert/ponytail
⭐ 154,245 stars
#2 · Panniantong/Agent-Reach
⭐ 90,361 stars
#3 · pbakaus/impeccable
⭐ 75,840 stars
#4 · calesthio/OpenMontage
⭐ 62,848 stars
#5 · coreyhaines31/marketingskills
⭐ 52,790 stars
#6 · getsentry/sentry
⭐ 45,274 stars
#7 · earthtojake/text-to-cad
⭐ 16,670 stars
#8 · tester-army/e2e
⭐ 2,452 stars
Papers & Researchacademic
LLM papers from arXiv cs.CL and Hugging Face Daily Papers.
KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards
LLMs are increasingly applied to cybersecurity workflows, where they are expected to translate analysts' intent into tool invocations. However, existing evaluations focus on knowledge-based assessments or end-to-end agentic tasks, and do not directly measure LLMs'.
Hierarchical Continuous Diffusion Language Models
Discrete diffusion language models offer a compelling alternative to autoregressive generation for tasks demanding bidirectional reasoning and global constraint satisfaction. Yet they share a structural bottleneck: when decoding in parallel, each token is sampled independently.
Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair
Ablate a component of a language model, and other components often appear to adjust and compensate. This phenomenon, termed self-repair, has been observed repeatedly, but its mechanism remains unclear.
AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents
Coding agents solve repository-level software engineering tasks through long trajectories of code inspection, search, editing, and testing. As a task progresses, earlier exploration becomes stale, so managing context is more than avoiding overflow: an agent.
From Knowledge Access to Source Learning: Developing Source-Specific Competence
Large language model (LLM) agents increasingly rely on persistent external sources to solve sequences of knowledge-intensive tasks. Existing methods improve how source content is accessed and organized, while agent-memory systems preserve reusable knowledge from prior.
Keyword Harnesses Fail Open: A Cheap Diagnostic Ladder for Tool-Use Claims in Small Language Models
Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive in a matched-architecture pair of Spanish security language models and propose a ladder of strict, cheap diagnostics.
Argo-Bench: Evaluating Data Agents on Enterprise-Scale Workflows
Real-world enterprise data science and analytics workflows require reasoning across dozens of tables, performing statistical analyses, and acting on the results. Established text-to-SQL benchmarks evaluate query generation alone, and audits have found their answer keys.
Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes
On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models.
Scalable, Transferable Meta-network for Data Selection Requires a Different Loss (and Why the Obvious Choice is Problematic)
Data selection is critical for training large language models on massive and heterogeneous corpora. Meta-learning for Training-data Selection offers a principled alternative to heuristic scoring by learning data weights from a target validation objective, but.
LLM2Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them
Jev-style decision models return categorical probability distributions over predefined options without generating free-form text, enabling software systems to act on their outputs directly. In this work, we investigate the extent to which general-purpose LLMs already.
Typological Alignment of Stack-Based Language Models on Mildly Context-Sensitive Artificial Languages
Some properties of languages, e.g., subject-object-verb (SOV) word order, are more prevalent than others among the thousands of attested natural languages (NLs). Such typological commonality is often attributed to learning biases.
CARM: Cancellation-Aware Response Masking for LLM Reinforcement Learning
Recent years have witnessed the rapid adoption of reinforcement learning (RL) in large language model (LLM) post-training, with substantial gains in mathematical reasoning and code generation. In practical systems, however, policy updates and differences between.
Old Ideas, Novel Problems: The Instability of LLM-Based Novelty Evaluation
Automated ideation systems are often evaluated on the novelty of the ideas they produce, and that judgment is increasingly delegated to large language models. Such judges are typically built ad hoc and validated, if at.
Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization
The rapid growth of video-based social media has increased users' exposure to harmful content, creating a need for reliable automated video safety detection. Although recent Vision-Language Models (VLMs) show strong video understanding capabilities, existing harmful.
Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents
Large Language Model (LLM) agents now take part in organizational work, where many authors record decisions across documents over months. Because a revised decision arrives as a new document rather than an edit, answering a.
Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities
Byte-level byte-pair encoding (BBPE) tokenizers are attractive for multilingual large language models (LLMs) because they cover all Unicode text. In UTF-8-based BBPE, however, many scripts start from a higher fallback cost than English: when no.
Counterfactual Auditing of Bias in Open-Source Large Language Models for Clinical Triage
Emergency department (ED) triage is a high-stakes prioritization task in which demographic, socioeconomic, and system-context information may improperly influence acuity assignment. Although open-source large language models (LLMs) are increasingly considered for local and privacy-preserving clinical.
Latent JEPA: Abstract Future Prediction for Latent Reasoning in Chemistry
Large language models offer a promising foundation for chemical reasoning, bringing together chemical knowledge and multistep problem solving. Chemical intuition can provide an initial sense of plausible outcomes before the details of a solution are.
A rubric landscape for evaluating clinical reasoning in large language models: what exists, what is missing, and what needs to be combined
Exam-style accuracy does not establish whether large language models (LLMs) reason well over clinical records. We define clinical reasoning as integrating and updating evidence across time and sources to form, revise and justify a patient's.
Cross-Lingual Alignment for Decoder-Only Models using MoE Routers
Cross-lingual contrastive learning has been a core component of multilingual encoder training, but the ability to explicitly align representations is not possible in decoder-only LLMs because of varying multilingual tokenization. However, a growing amount of.
MoLE: Mixture of Latent Experts for Complementary Visual Reasoning
Latent visual reasoning equips vision--language models with continuous intermediate states that can process visual evidence without explicit textual reasoning traces or repeated image operations. However, existing methods often allow multiple latent tokens to access the.
Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It
A small LoRA is enough to stop transformers from quitting a chain of thought early.
Leaderboardsrankings
Vellum's published top 5 per benchmark.
Humanity's Last Exam
| 1 | Claude Opus 5.5 | 67.7 |
| 2 | Claude Fable 5.1 | 65 |
| 3 | Claude Mythos 5.1 | 65 |
| 4 | Claude Opus 5 | 64.7 |
| 5 | Claude Mythos 5 | 64.5 |
GPQA Diamond
| 1 | Claude Sonnet 5 | 96.2 |
| 2 | GPT-6 Astra | 96 |
| 3 | Claude 3 Opus | 95.4 |
| 4 | GPT-5.6 Sol | 94.6 |
| 5 | Gemini 3.1 Pro | 94.3 |
SWE-Bench
| 1 | GPT-5.6 Sol | 96.2 |
| 2 | Claude Mythos 5 | 95.5 |
| 3 | Claude Fable 5 | 95 |
| 4 | GPT-5.6 Luna | 93 |
| 5 | Claude Opus 4.8 | 88.6 |
News & Analysisindustry
Newsletters, lab blogs, researcher Substacks, HN.
Simon Willison
We're going to need default hard budget caps on pretty much everything
2026-10-03
Always-on agents need default hard spend caps. A soft warning is not a control once a loop can keep buying tokens.
Hacker News
Getting the most out of Opus 5.5 in Claude and Claude Code
2026-10-03
221 points · 152 comments (221 pts · 152 comments)
AI as Normal Tech
A big-tent or small-tent AI safety movement?
2026-10-01
In recent weeks, two narratives about AI safety have emerged: either AI existential risk is real and imminent, or AI leaders’ and whistleblowers’ claims to that effect are insincere — a “psyop” or hype or.
Hacker News
GPT-Synopsys: Frontier Intelligence to Revolutionize Chip Design
2026-10-01
189 points · 111 comments (189 pts · 111 comments)
Latent Space
Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week
2026-09-30
Three months ago Dwarkesh, who has been posting incredible blogs and episodes about RL, posted a framing question for his video essay on RLVR which upset a lot of Computer Use folks:We are no strangers.
DeepMind
Gemini 4 Argon: our next era of frontier intelligence
2026-09-30
Gemini 4 Argon is DeepMind's new flagship, with a 1 million token output and first access for cyber defenders.
DeepMind
Introducing SynthID Bio
2026-09-30
SynthID Bio is a proof of concept for watermarking AI-generated proteins without breaking their function.
OpenAI
Disrupting a coordinated model-distillation campaign
2026-09-30
OpenAI says it broke a distillation campaign that was extracting protected reasoning, and ties a cluster of it to people associated with Moonshot.
HuggingFace Blog
Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning
2026-09-30
Hugging Face opens a leaderboard for multilingual text-to-speech and voice cloning.
HuggingFace Blog
NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction
2026-09-29
HuggingFace Blog
Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents
2026-09-29
Ahead of AI
Language Models for Text Classification: From Bag-of-Words to Jev
2026-09-29
The recently released Jev AI model has been quite a cultural phenomenon in technical communities in the past 2 weeks.While Jev aims to classify things, it’s easy to dismiss Jev as “just a classifier,” and.
OpenAI
DevDay 2026 Recap
2026-09-29
DevDay 2026 shipped GPT-6.1 Sol, Dots, and a stack of agent, decisions, and marketplace APIs. ChatGPT is at 1.2 billion weekly users.
OpenAI
Introducing GPT-6.1 Sol
2026-09-29
GPT-6.1 Sol prices near-Astra coding and computer-use at one-fifth of Astra's API token rates.
Latent Space
[AINews] AMD buys World Labs for $8.2B, as Atlas solves sparse reconstruction problem for robotics, design and more
2026-09-29
AMD is buying Fei-Fei Li's World Labs for $8.2 billion. The lab's Atlas work targets sparse 3D reconstruction.
OpenAI
Introducing dots
2026-09-29
Dots are OpenAI's always-on assistants that keep working a project after the chat ends, with the user still able to stop them.
Simon Willison
Claude Sonnet 5.5
2026-09-28
Claude Sonnet 5.5 is priced like the previous Sonnet and Anthropic says it is 30% faster and up to 30% cheaper on most work.
OpenAI
How we will do better for Australia
2026-09-28
OpenAI apologizes for agents hitting Australian government sites and says it is adding safeguards.
OpenAI
Towards safety cases for frontier AI training
2026-09-28
OpenAI publishes early guidelines for safety cases on frontier training: technical safeguards, operations, and misalignment investigations.
Mistral
Hallo, Deutschland!
2026-09-28
Mistral is opening a Munich hub for physics and industrial AI research.
Import AI
Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop
2026-09-28
Import AI 474 covers a shared 'platonic' representation hypothesis, TPUs in orbit, and Zhipu starting an outer RSI loop.
AI as Normal Tech
AI existential risk probabilities are (still) too unreliable to inform policy
2026-09-28
Over two years ago we wrote a detailed deconstruction of p(doom) and argued that its primary effect is to launder vague intuitions and fears through a facade of quantification. The essay has fresh relevance today.
Agent curation log
2960 items reviewed161 kept2799 dropped