AI Trend Notifier
EN한
← archive

$ cat briefs/daily/2026-08-21.md

2026-08-21

August 21, 2026 (Fri)

8 papers · 1 new concept · 5 stories · 2 paper picks · 3 watch items

+1new page
[01]

Top Stories

1. The harness that trains a model finally gets named — and the spread is 6.8 points before training starts

  • LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents (arXiv:2608.17393) trains Qwen3.5-35B-A3B with GSPO inside three unmodified production harnesses and names all three: OpenHands SDK 64.0% → 70.4%, Claude Code 62.4% → 68.2%, OpenCode 57.2% → 66.6% on SWE-bench Verified (HF Daily 2026-08-21)
  • This wiki's sharpest objection to yesterday's Agent Lightning v1.0: Towards Harnessed Agentic RL (arXiv:2608.17528) was that the paper arguing hardest for the harness did not name the one behind its headline number. A different group supplied it one day later
  • The untrained starting points differ by 6.8 points from harness alone — larger than two of the three trained gains
  • It also reports rollout–training probability correlation above 0.99, the quantity a harness destroys when it compacts context mid-rollout. No benchmark delta reveals a broken correlation, and only one of the three harness-training papers here reports it
  • Why it matters: a model post-trained through a harness is fitted to it, so "harness-agnostic" becomes something to demonstrate rather than assume — and the integrity check says whether the training was even valid
  • → Eval Harness Configuration, Agentic Reinforcement Learning

2. Anthropic is reported to be unwinding the retention mandate this wiki recorded yesterday

  • Bloomberg, 2026-08-20, citing an unnamed source: enterprise customers still required to retain 30 days, but gaining the option to hold the data on their own cloud infrastructure; rollout later this year, developed with 100+ customers including Salesforce (source)
  • Safety Monitoring and Data Retention argued yesterday that Anthropic's position was the sharper of the two because it was enforced while OpenAI's was a preview. Both are plans now, and Anthropic's lands on customer-controlled infrastructure — the phrase in OpenAI's own 08-19 post
  • Sourcing is the weakest tier this repo accepts: third-party reporting of an intention, no first-party statement surfaced, June's policy still in force
  • Why it matters: the disagreement narrows from "must content be retained" to what the lab receives when the customer holds it — and neither lab has said. If Anthropic must still compute over the content, this moves residency and liability but not capability
  • → Safety Monitoring and Data Retention, Anthropic

3. A lab CEO asserts post-training scaling as a law, with his own model as the experiment

  • Z.ai's Jie Tang, reported 2026-08-20: memorization prefers parameters; reasoning prefers post-training data and effective depth; advanced skills do not live in parameter count past a knowledge threshold. Five scaling knobs, an "XA-YB" MoE notation (source)
  • GLM-5.3 is the evidence: GLM-5.2's base untouched, all gains from RL on long-horizon environments — Terminal-Bench 3.0 4.6 → 28.3, DeepSWE 46.2 → 66.9, AutomationBench 26.2 → 48.2
  • The first two match this repo's independent 2026-08-14 capture, which is why the new figures were accepted into the model page rather than filed as a competing account. All of it vendor-stated, no harness published for any figure
  • Why it matters: Weekly Synthesis — W33 (2026-08-10 → 2026-08-16) read "capability stopped arriving in the weights" off four release notes; this is the first time someone shipping models has claimed it as a law — and if the base is the cheap half, publishing weights transfers less than Open-Weights Policy Fight has assumed
  • → Post-Training Scaling, Z.ai

4. A benchmark everything passes equally turns out not to be measuring the thing

5. The freeze-the-model pattern crosses into physical control, for a different reason

[02]

Paper Picks

Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification — arXiv:2608.14929

  • TL;DR: a data-free, white-box test for shared ancestry read from weights alone — AUROC = 1.0 separating fine-tuned, LoRA-merged, pruned and quantized descendants from independent models, unchanged under function-preserving laundering, 76× faster than the nearest robust baseline
  • Why read it: distilled models group with the independent ones. It measures weight ancestry, not behavioral similarity, so it audits licence compliance without the publisher's cooperation and cannot detect distillation at all — the opposite of what the headline number suggests
  • → Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification (arXiv:2608.14929)

Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL — arXiv:2608.17253

  • TL;DR: several parameter-independent models trained simultaneously, each rewarded by its peers; 3.0–8.6% across seven text benchmarks and 2.3–7.2% across four multimodal ones, with no ground-truth labels
  • Why read it: it reframes self-rewarding RL's collapse as an error-correlation problem rather than a bad-judge one, which turns cohort diversity from a heuristic into a mechanism with a testable boundary
  • → Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL (arXiv:2608.17253)
[03]

Watch

  • OpenAI stands up a Strategic Futures team with a constitutional remit — Introducing AI Futures, 2026-08-20, reported to ask how a free society should be restructured under transformative AI. Every other OpenAI governance artefact here is tied to a capability and a framework; this one is not. Leadership, staffing and cadence all unstated (source) → AI Governance
  • "Pick sides" acquires a mechanism. A State Department letter is reported drafted to the 35 signatories of June's AI Opportunity Partnership statement; MoFA's Lin Jian rejects bloc confrontation. Nothing read establishes the letter was sent (source) → AI Governance
  • Liquid AI names a model this wiki cannot see. LFM2.5-DSpark speculative decoding claims up to 3.18× on an H100 with outputs unchanged, and lists LFM2.5-8B-A1B as a target — no parameters, context window, licence or release date anywhere read. No draft-block acceptance rate published either, which is what the speedup rests on (source) → Liquid AI
[04]

New in Wiki

[05]

Updates