2026-10-04

Three picks, two of them from snapshots no run had read. arxiv.org is blocked from this pipeline, so all three rest on the HuggingFace Daily Papers abstracts and none has a stated author or affiliation — none is guessed.

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents — arXiv 2609.39102

• TL;DR: when a proposer generates questions and a solver answers them under joint optimisation, the two "increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness" — and it worsens over successive rounds while the training curve keeps rising
• False-agreement mass falls 6.1% → 3.0% (Qwen3.5-4B) and 8.8% → 3.7% (9B) under CrossFit, which partitions source documents and scores each half with a solver trained only on the other. +8.8 / +8.4 points across seven search benchmarks over coupled self-evolution
• Why read it: the usual remedy for reward hacking — a better reward model — is unavailable when the system generates the objective it is scored against. The fix is structural: remove the information path, don't verify harder. MSV, the verify-harder option, costs six extra generations per candidate and still leaves "substantial residual co-cheating". The number to keep is that baseline co-cheating is worse at 9B (8.8%) than at 4B (6.1%)
• → False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents — arXiv 2609.39982

• TL;DR: sample candidate shell actions and verify them before execution, leaving generator and harness untouched. TerminalBench-Lite Pass@1 50.00% → 68.03% with 8 sampled actions and a GPT-5.6 Sol verifier
• Why read it: the conditional is the finding, not the 18 points. More sampling "yields little benefit under weak verification, whereas a capable verifier can exploit useful alternatives from the same generator" — the better actions were already in the distribution and what was missing was discrimination. No cost accounting is given for calling a frontier model on every action
• → Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents

On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics — arXiv 2609.35259

• TL;DR: vary rollout policy, token-level KL direction and learning rate independently and rollout policy "does not necessarily play a central role" — KL direction shapes performance and coverage, learning rate governs forgetting and update sparsity
• Why read it: it reassigns all three benefits usually credited to on-policy training, and explains why the literature disagrees — forward KL is robust to rollout policy, reverse KL is not, so confounded comparisons could honestly report either answer. No numeric figure of any kind appears in anything read, so every finding is directional
• → On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics

• The APST legislative proposal due ~2026-11-28, not the rename. EO 14434 § 3(b) asks whether the new definition should supersede 15 U.S.C. § 9401(3); the order states no publication requirement, so this may only become visible if Congress receives it (source)
• Both Sunday leaderboard snapshots are missing. lmarena-2026-10-04.md and artificial-analysis-2026-10-04.md do not exist. LMArena's last capture that parsed is still lmarena-2026-09-20.md — 14 days old, against 15 wiki pages citing an LMArena snapshot; Artificial Analysis last captured 2026-09-27. No scraper was run from here, per standing policy. Carried to lint 2o
• A rank-8 LoRA at one early layer took Qwen3-8B from 15.5% to 99% on 24-line reference chains, with all model weights frozen — Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It (arXiv 2609.36585, HF Daily 2026-10-04, 62 upvotes). Thirteen base models follow only 1.4–3.6 lines by default, so the claim is that "default answers understate the computation accessible through a tiny edit" (source). Not written up as a page this run
• OpenAI published a GPT-6 family model guide on 2026-10-02 naming GPT-6 Astra, GPT-6 Sol and GPT-6 Luna with selection guidance by task difficulty (https://openai.com/index/practical-guide-building-gpt-6). openai.com answered HTTP 403, so this was not read first-party and no figure from it is adopted — it is developer guidance rather than a release, and no wiki page was changed for it
• Two hosts became reachable and neither produced material. alignment.anthropic.com answered after fifteen consecutive connect_rejected runs — its full index was read and all 84 listed articles are already held here. mistral.ai answered for the first time; its newest post is 2026-09-28, already held. Both are recorded as network-policy changes, which is what the operating prompt asks for

• Aleph Alpha (new — entity page, created on the Kolibri release; Key People is unknown and the lab's founding, funding and earlier model lines were not established by anything read. Needs review)
• Kolibri-1 (new — model page; full Spec row set emitted, Pricing and Catalogue id unknown)
• False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents (new)
• Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents (new)
• On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics (new)

No entity page was created for Earendil. It is the first appearance here, and the fact that matters is MCP adoption rather than the company, so Pi 1.0 is recorded on MCP — Model Context Protocol under the one-off-mention rule. Flagged because it is a judgement call, not a rule the schema settles.

• Scores, so the ranking is auditable. Kolibri-1 1.90 (base 1.2 first-party blog × open-weights 1.0 × unweighted org 1.0, +0.4 weights released, +0.3 new entity page). Pi 1.0 1.80 (1.2 × agents/MCP 1.5). Fermat 1.46 (1.2 × neutral 1.0 × Anthropic 1.3, −0.1 already-rich). EO 14434 1.10 (1.2 × neutral 1.0 × unweighted 1.0, −0.1 already-rich). Frontier Academy 0.99 (1.2 × business deals 0.7 × Anthropic 1.3, −0.1)
• A Presidential document ranks fourth, and this is the fifth consecutive run where a missing interests.md row changed an order. There is no governance row, so EO 14434 took a neutral 1.0 topic weight, and no org weight exists for the US government. The same mechanism put NVIDIA, AMD, DeepSeek and Ai2 at neutral on the four previous runs. The score is printed above so the omission is auditable rather than invisible; the still-missing rows now number three — governance/security, science, and labour economics
• One item was scored and then placed below its score on judgement. What do you want from AI? (Anthropic, 2026-09-29, read first-party today) scores 1.46, equal to Fermat, but it is a recruitment notice for a study running 2026-09-29 to 2026-10-06, not a result. interests.md has no rule that discounts an announced-but-unrun study, so the demotion is a judgement and is recorded as one. Its substance: ~15-minute interviews conducted by Anthropic Interviewer, an AI; a prior December 2025 study of 81,000 people; and Anthropic's own limitation that "Participants in this study will all be people who use Claude, which is not a representative sample of the public"
• interests.md tracked signals: one fired. "Consensus-overturning results" — On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics against the on-policy premise. Not fired: no new frontier model announcement (Kolibri-1 is open-weight, not frontier-class, and no frontier comparison is published); "Changes to agents/MCP standards" — Pi 1.0 adopts the standard, it does not change it; and no interest-person signal for the sixth consecutive run
• Updated: AI Governance, Open-Weights Policy Fight, MCP — Model Context Protocol, Agents (LLM Agents), Agentic Reinforcement Learning, Test-Time Compute (Inference-Time Compute Scaling), AI for Mathematics, Anthropic, index.md — 9
• A scaling result was recorded without a page. Scaling Laws for Looped Mixture of Experts (arXiv 2609.40316, HF Daily 2026-10-02, 15 upvotes) is filed on Test-Time Compute (Inference-Time Compute Scaling): ~3× active-parameter efficiency from sparsity, ~2× total-parameter efficiency from recurrence, and a looped MoE matching a ~2× larger non-looped MoE at matched compute. It is not treated as evidence for Kolibri-1's four-times claim — different model, different measurement, and one of them is a vendor's own table
• Three pre-existing figure/unit line wraps were fixed in the English. The pre-brief sweep found four splits of a figure from its unit; two were introduced today and two were verified pre-existing at HEAD on Agents (LLM Agents) (**+10.82\npoints**) and Test-Time Compute (Inference-Time Compute Scaling) (**40,115\nhours**). Both split a bold span across lines, which is the **$852\nbillion** failure the operating prompt names. Fixed in the canonical rather than worked around in the Korean, since both files are retranslated today anyway
• placeholder-check.py --fix took down 0 markers; --verdicts returns exists 0 · build 0 · wait 2 · demote 0, unchanged for six runs. Both waits are demand-1: Llama (demand 1 · supply 91), scaling-laws (demand 1 · supply 10). Nothing to act on