LLM papers, model releases, leaderboards, and infrastructure — arXiv, HuggingFace, Latent Space, Import AI, lab blogs, and live Vellum leaderboards.
Cross-benchmark aggregation across public leaderboards (Vellum, Artificial Analysis, EQ-Bench, LiveBench, LMSys). Sources that stop updating are auto-excluded — see below.
31 models ranked across 15 benchmarks — aggregated from multiple sources. Top composite: Claude Fable 5, Claude Opus 5, Kimi K3.
Data as of 2026-08-17 · live sources: Artificial Analysis, LMArena, Vellum.
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 |
|---|---|---|---|---|---|
| 1 | Claude Fable 5· | 25–0–1 | 78th | 12 | Artificial Analysis, LMArena, Vellum |
| 2 | Claude Opus 5· | 24–1–0 | 87th | 8 | Artificial Analysis, Vellum |
| T3 | Kimi K3▲1 | 22–2–1 | 82nd | 8 | Artificial Analysis, Vellum |
| T3 | GPT-5.6 Sol▼1 | 21–1–3 | 78th | 12 | Artificial Analysis, Vellum |
| T5 | Gemini 3.7 Flashnew | 19–3–3 | 79th | 8 | Artificial Analysis |
| T5 | Grok 4.6new | 20–4–1 | 78th | 7 | Artificial Analysis |
| T5 | Muse Spark 1.2▼2 | 19–5–1 | 75th | 8 | Artificial Analysis, LMArena |
| T8 | Qwen3.8 Max▼1 | 18–7–0 | 68th | 9 | Artificial Analysis, LMArena |
| T8 | GPT-5.6 Terra▼1 | 17–8–0 | 65th | 11 | Artificial Analysis |
| 10 | GLM 5.2▲3 | 15–10–0 | 67th | 10 | Artificial Analysis, Vellum |
Darker = higher score within each benchmark column. Columns use different scales, so compare within a column, not across.
| Model | AA Intelligence Index | GPQA Diamond | Humanity's Last Exam | SciCode | AA Agentic Index | AA Long Context Reasoni… | LMArena Elo | MMMU-Pro | IFBench | Tau2 |
|---|---|---|---|---|---|---|---|---|---|---|
| Claude Fable 5 | 62 | 93 | 64 | 60 | 57 | 77 | 1506 | — | 64 | 98 |
| Claude Opus 5 | 63 | 93 | 65 | 56 | 59 | 76 | — | 85 | — | — |
| Kimi K3 | 60 | 94 | 56 | 59 | 54 | 83 | — | 80 | — | — |
| GPT-5.6 Sol | 61 | 95 | 50 | 56 | 58 | 78 | — | 83 | 73 | 85 |
| Gemini 3.7 Flash | 56 | 94 | 48 | 57 | 45 | 80 | — | 86 | — | — |
| Grok 4.6 | 61 | 95 | 43 | 54 | 59 | 75 | — | — | — | — |
| Muse Spark 1.2 | 57 | 90 | 46 | 56 | 49 | 83 | 1498 | — | — | — |
| Qwen3.8 Max | 58 | 93 | 43 | 53 | 58 | 74 | 1491 | 82 | — | — |
| GPT-5.6 Terra | 57 | 92 | 43 | 54 | 50 | 80 | — | 81 | 71 | 86 |
| GLM 5.2 | 53 | 90 | 55 | 50 | 46 | 77 | — | — | 73 | 99 |
| DeepSeek V4 Pro 0813 | 53 | 93 | 41 | 49 | 50 | 75 | — | — | — | — |
| GPT-5.6 Luna | 52 | 91 | 40 | 52 | 47 | 78 | — | 79 | — | — |
LLM papers from arXiv cs.CL and Hugging Face Daily Papers.
Structure-property relationships are foundational to biology, chemistry and materials science, where function, reactivity and physical response emerge from spatial, chemical and periodic organization. Mechanistically explaining these relationships requires interpreting structural evidence through scientific principles…
Limited memory language models (LMLMs) externalize factual knowledge during pretraining to a knowledge base (KB), rather than memorizing it in their weights. During generation, the model then fetches knowledge from the KB as needed.
An arXiv paper on LLM agent architectures, training, or evaluation.
Reinforcement learning from verifiable rewards (e.g. GRPO) is the engine behind today's reasoning models, yet it grades only the final answer.
An arXiv paper advancing the GRPO family of RL post-training algorithms — likely an improvement on a baseline or a new variant.
Large language models hallucinate most about entities they have never seen. We ask whether a model's activations betray entity familiarity before a single answer token is generated, and whether that signal predicts the factual reliability.
Large language models increasingly \emph{understand} dialectal English, yet still \emph{produce} only standard, US-leaning English, leaving dialectal generation, the harder half of the problem, largely unaddressed. We introduce \textbf{DiaLLM}, which continually pretrains three open-weight language model.
Does RL post-training merely amplify primitive skills already latent in a base model, or can it compose primitive skills into new higher-level strategies? We study this question in a fully observable rewrite-grammar environment where the.
Reliable confidence estimation is essential for deploying large language models (LLMs) in confidence-aware systems, where downstream decisions such as retrieval, tool use, and adaptive computation depend on accurately estimating answer reliability. Existing approaches, however, largely.
One-shot pruning methods like Wanda and SparseGPT apply the same sparsity ratio to every layer of a transformer, ignoring known variation in layer importance. We propose PALS (Percentile-Aware Layerwise Sparsity), which adjusts per-layer sparsity based.
An arXiv paper on LLM agent architectures, training, or evaluation.
FFT-based spectral preprocessing of learned query-key (Q/K) projections substantially improves transformer attention on character-level language modelling. On TinyShakespeare: a fixed random spectral filter achieves val=1.031 (Delta=+0.443); a single learned frequency at paragraph scale achieves val=0.608.
An arXiv paper on LLM agent architectures, training, or evaluation.
Fine-tuning large language models (LLMs) for e-commerce attribute extraction requires labeled data representative across thousands of product types, attributes, and multiple languages. This combinatorial scale translates to millions of annotations, rendering human labeling prohibitively costly.
An arXiv paper on LLM agent architectures, training, or evaluation.
Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an entire block in one pass, but their position-wise predictions.
Automatic prosodic segmentation identifies boundaries between speech units from acoustic and linguistic evidence. Although recent deep learning approaches have produced strong results for English, automatic segmentation for Brazilian Portuguese (BP) still relies mostly on rule-based.
Large Language Models (LLMs) store factual knowledge and domain-specific patterns implicitly in dense Transformer parameters, making knowledge expansion costly through pretraining, fine-tuning, retrieval augmentation, or longer contexts. Engram-style memory offers a compact hidden-state injection pathway.
An arXiv paper on LLM agent architectures, training, or evaluation.
This paper reports an empirical study evaluating the relevance of several RAG metrics. The experiment is based on a question-answering dataset created by human annotators from business data.
Harmful online communication often contains slang, coded terms, abbreviations, and community-specific expressions, which make messages difficult to interpret. This paper presents an exploratory study of interpretation difficulty in Discord chats related to cybercrime.
HF Daily Papers feature on a robotics paper.
HF Daily Papers feature on a video / generation LLM paper.
HF Daily Papers feature on a reinforcement-learning LLM paper.
HF Daily Papers feature on a robotics paper.
Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
HF Daily Papers feature on an LLM agent paper.
HF Daily Papers feature on an LLM agent paper.
HF Daily Papers feature on a reinforcement-learning LLM paper.
HF Daily Papers feature on an LLM agent paper.
Vellum per-benchmark top-5 rankings.
HuggingFace trending models.
5,572 downloads · 600 likes
1,875,602 downloads · 1921 likes
362,300 downloads · 3720 likes · $3.00/M out
23,112 downloads · 427 likes
1,246,042 downloads · 1900 likes
16,374 downloads · 328 likes
4,647 downloads · 194 likes
821 likes
258 downloads · 176 likes
2,189 downloads · 172 likes
Newsletters, lab blogs, researcher Substacks, HN.
Simon Willison's coverage of a GPT release or analysis.
786 points · 578 comments
Introducing Muse Spark 1.1 Following Muse Spark in April, here's Muse Spark 1.1 - the first Spark model to offer an API. Meta claim significant improvements in agentic tool calling and computer use.
Simon Willison's coverage of an LLM topic.
Simon Willison's coverage of an LLM topic.
Inviting hard questions
Ben Bernanke appointed to Anthropic’s Long-Term Benefit Trust
Introducing a way to reflect on how you use Claude
OpenAI releases GPT-5.6, the next step in the GPT-5 frontier family.
Studio provides a system of record for AI prompts and skills—versioned, owned, and traceable. Iterate fast, ship with control, and ensure consistent AI behavior.
ChatGPT Work is an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.
OpenAI GPT-5.5 release or analysis.
OpenAI releases GPT-5.6, the next step in the GPT-5 frontier family.
As GPT 5.6 is confirmed to launch tomorrow, today is pretty much the last day anyone will be excited about a GPT 5.5 equivalent model launch, and that is exactly what SpaceXAI did:The new Grok.
Rewriting Bun in Rust Jarred Sumner has been promising this blog post (since May 9th) about his Zig to Rust rewrite of Bun for significantly longer than it took him to finish the rewrite. Honestly.