BlackLeafwatch the watchmen
AI reasoning research, digest window 2026-W33: one open-weight model release plus three arXiv preprints on proof generation, test-time-compute scaling, and diffusion-based world modeling

Reasoning: the headlines outrun the science

Summary

An open-weight model out of Xiaohongshu, dots3-note (preview) -- 280B parameters, 16B active, 512K context, Apache-2.0 -- shipped this cycle to headlines about a near-perfect math-olympiad run. The harder research underneath is more sobering. A fresh theoretical-CS benchmark finds the best frontier model can prove only 204 of 300 research-level theorems. A new survey argues the field hasn't even agreed how to measure 'thinking longer.' And a 399M-parameter diffusion transformer is trying to fuse world-modeling into robot control with no VLM backbone at all.

By Marcus Aurelius · August 14, 2026

This cycle's headline was an open-weight model's math-olympiad run. The research underneath is more sober. On TCS-Bench, a new benchmark of 300 research-level proof tasks drawn from STOC, FOCS, and SODA, the best frontier model -- GPT 5.6 Pro -- proves just 204 of them (68%). Competition-math trophies are not the same as writing real theorems.

The release that made the headlines

Xiaohongshu's dots3-note (preview) is a 280B-parameter Mixture-of-Experts model with 16B active per token, a 512K context, and an Apache-2.0 license -- the first open-weights drop of the dots3 line. It arrived to reports of a top math-olympiad score. But on the model card those benchmarks are chart images, with the full report still "coming soon" -- the sensational number is out ahead of any released technical write-up.

The research underneath

Research-level proofs: how far the frontier gets
Accuracy on TCS-Bench, 300 theorem-proving tasks from STOC / FOCS / SODA
GPT 5.6 Pro
68%
Gemini 3.1 DeepThink
52%
Opus 5
32.8%
Gemini 3.1 Pro
30.3%
Source: TCS-Bench, arXiv:2608.09538, Table 1 (2026)
View data as table
GPT 5.6 Pro68%strongest: 204 of 300 proofs
Gemini 3.1 DeepThink52%next best
Opus 532.8%hit its 128K token budget on 162 of 300 tasks
Gemini 3.1 Pro30.3%base model, no agentic harness
Best model on TCS-Bench
68%
GPT 5.6 Pro — 204 of 300 research-level proofs
dots3-note open weights
280B / 16B
MoE total / active; 512K context; Apache-2.0
WorldDiT, no VLM backbone
399M params
94.9% mean across four LIBERO suites
  • Open weights, big claim. Xiaohongshu open-sourced dots3-note (preview) -- a 280B-parameter MoE, 16B active, 512K context, Apache-2.0 -- to headlines about a near-perfect math-olympiad run. The catch: the card's benchmarks are chart images and the full report is still "coming soon," so the number is ahead of any released science.
  • Research-level proofs stay hard. TCS-Bench grades models on 300 self-contained proof tasks from STOC/FOCS/SODA; the best, GPT 5.6 Pro, solves 204 of 300 (68%), and its automated grader agrees with human experts more than 90% of the time. Olympiad wins don't transfer to writing real proofs.
  • We can't agree how to measure "thinking longer." A survey, Test-Time Scaling in Reasoning LLMs, argues the field lumps distinct inference methods under one scalar "budget" and reports accuracy without the protocol that produced it -- and releases over 2 billion reasoning traces to make the comparisons reproducible.
  • World-modeling without a VLM. WorldDiT fuses action generation and future-frame prediction in a single diffusion transformer with no pretrained VLM backbone, hitting 94.9% mean success across four LIBERO suites at just 399M parameters -- on the parameter-vs-success Pareto frontier among 24 methods.
WorldDiT: a 399M-parameter policy, no VLM backbone
Success rate across the four LIBERO simulation suites (mean 94.9%)
Spatial
98%
Object
97%
Goal
92.8%
Long
91.8%
Source: WorldDiT, arXiv:2607.23909 (2026)
View data as table
Spatial98%
Object97%
Goal92.8%
Long91.8%hardest suite

Digest window 2026-W33; every item's facts trace to the primary release or preprint, not a news write-up. The dots3-note math-olympiad result is not stated in verifiable text on the model card (its benchmarks are chart images with the report "coming soon"), so this piece attributes that claim to the headlines and does not treat it as a verified figure -- the confirmed dots3 facts are its specs. TCS-Bench scores are graded by the paper's own automated verifier (self-reported >90% agreement with human experts), not by independent human grading. WorldDiT's 94.9% mean is the authors' own figure, which they caution is not an unbiased test estimate because staged checkpoint selection reused evaluation episodes.

Sources(5) ▾
  • dots studio (Xiaohongshu / rednote-hilab), dots3-note (preview) — model card (2026-08)The issuer-hosted Hugging Face model card for the first open-source release of the dots3 series. Source for the model's total/active parameter counts, context length, and Apache-2.0 license, and for the fact that its benchmark results are published on the card only as chart images (bench_en1.png / bench_en2.png) with a 'Full Report (coming soon)' note — i.e. no textual technical report yet backing the widely reported math-olympiad claim. Fetched directly (rendered page and raw README). huggingface.co
  • dots studio (Xiaohongshu), studio-dots-ai/dots3-note-prev — repository (2026-08)The developer's GitHub repository for the same preview release. Corroborates the 280B-total / 16B-active MoE parameter counts, 512K context length, and Apache-2.0 license stated on the Hugging Face card. Fetched directly. github.com
  • Cohen-Addad, Paparas, van Wijland, Springer, et al. (arXiv:2608.09538), TCS-BENCH: Benchmarking State-of-the-Art Generative AI Theoretical Computer Science Research Ability (2026-08-10)arXiv preprint introducing TCS-Bench, a benchmark of research-level theoretical-computer-science proof tasks drawn from STOC, FOCS, and SODA. Source for the 300-task size, the per-model accuracy table (Opus 5, Gemini 3.1 Pro/DeepThink, GPT 5.6 Pro), the strongest-model result (GPT 5.6 Pro, 204/300, 68%), the reference-verifier accuracy (>90% vs human experts on 100 labeled proofs), and the Opus-5 token-budget caveat. Abstract and full PDF fetched and read via pdftotext. arxiv.org
  • Hariri, Chen, Shahini, Singh, et al. (arXiv:2608.04001), Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility (2026-08-04)arXiv preprint systematizing 'test-time scaling' into three inference regimes (single-trajectory sequential, leaf-level, prefix-level) and arguing that treating them as interchangeable under one scalar 'budget,' or reporting accuracy without the inference protocol that produced it, makes results hard to compare. Source for the three-regime taxonomy and the 'over 2 billion full reasoning traces' release. Abstract fetched and read. arxiv.org
  • Wang, Praveen, Roy, Villagra (arXiv:2607.23909), WorldDiT: A Unified Diffusion Architecture for World and Action Modeling (2026-07-27)arXiv preprint introducing WorldDiT, a diffusion transformer that couples continuous-action generation with visual world modeling and reaches its results without a large pretrained VLM action backbone. Source for the 399.084M total / 135.107M trainable parameter counts, the four-suite LIBERO success rates (98.0 / 97.0 / 92.8 / 91.8, mean 94.9%), the Pareto-frontier claim across 24 methods, and the authors' caveat that the 94.9% is not an unbiased test estimate. Abstract and full PDF fetched and read via pdftotext. arxiv.org
Weekly digest: the most-read systems, in brief. Mondays.

Comments

Always open. Logged-in readers can annotate paragraphs in place.

Loading comments…
or log in to comment under your account