2026-09-10

Today's HuggingFace snapshot turned over completely — 25 entries, 25 new after arXiv-id dedup, the first full turnover this repo has recorded.

What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets — arXiv:2609.05663 (1.85)

• TL;DR: two production fleets, real money, 7.5M single-model invocations and 231,638 multi-tool turns producing 14,596 fills — and the model is the part that does not move the outcome. A risk slider explains leverage at +0.425 per level, agent fixed effects absorb 60% of variance, a leaderboard render boundary routes selection at a 1.75× regression discontinuity, and on 416 replayed production scenarios frontier models are statistically indistinguishable at this horizon
• Sizing ignores volatility outright: median leverage 5.0× in every volatility sextile, with one posture cell holding 11% of the book and 62% of liquidations
• Why read it: it is Eval Harness Configuration's thesis measured where the harness is a product surface instead of an eval rig — and the one figure to carry carefully is the indistinguishability result, which is scoped to one task family at one horizon and is not evidence that frontier models are interchangeable
• → What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets

NeoHorse-1 — arXiv:2609.08183 — and On-Policy Reverse Distillation — arXiv:2609.08798 (1.85 · 1.59)

• Published as a pair because they answer the same question from opposite ends: where does a post-training signal come from when there is no stronger model to copy? NeoHorse-1 takes it from the routing layer — each turn's predicted capability demand, service tier and interaction recorded, validated, and replayed as training examples that keep interleaved reasoning, tool calls and harness context, with evaluation feedback setting the next mixture. Macro-average 58.94 → 64.87 at 4B and 65.60 → 69.04 at 9B over eleven unnamed benchmarks, putting post-trained 4B within 0.73 of the 9B base
• OPRD takes it from a weaker teacher and refuses to imitate it: read the teacher's policy shift against its own reference policy on the student's rollouts, then amplify only the verifier-supported component of the student's gradient along that direction — which preserves the stationary points of policy optimization, so the weak teacher cannot impose its ceiling by construction
• Why read them: OPRD is the third paper in twelve days moving authority from teacher to verifier, after One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation (asymmetry as a bias to manage) and FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience (verifier as a sign). It makes the verifier an admission gate, and none of the three cites either of the others. NeoHorse-1's "recursive" covers one closed loop — it calls itself an initial prototype — which is precisely the repetition question the OPSD review says degrades
• → NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness · Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

Steering Geometry: Validating Human Value Geometry in LLM Steering Space — arXiv:2609.06289 (1.59)

• TL;DR: two families of activation steering score the same and only one is steering what it claims to. Judged against Schwartz's Theory of Basic Human Values over a 26K-sample, 20-value benchmark, distribution-driven methods (CAA, SphericalSteer, ODESteer) recover the theory's predicted topology at Spearman ρ up to 0.51, p < 10⁻¹³; behavior-centric methods (COLD-Steer, BiPO) match on steering performance with little correlation to it
• Geometric fidelity rises with model scale and drops after instruction tuning — that is, drops on the models that get deployed, and the size of that drop is not published
• Why read it: it is the first result on Mechanistic Interpretability validated against a structure defined outside the model, which is the only reason the two families are separable at all. ρ = 0.51 is a ceiling and a moderate one; the honest reading is that one family recovers some of the predicted structure and the other recovers essentially none
• → Steering Geometry: Validating Human Value Geometry in LLM Steering Space

• Procedural knowledge as a graph rather than a transcript. Procedural Graphs: Self-Evolving Execution Structures for LLM Agents stores (procedure, relation, procedure) triplets, renders the subgraph around the agent's active node as guidance that biases without dictating, and evolves by contrasting failed against successful trajectories — committing only edits that preserve or improve held-out validation and keeping the rejected ones so the loop stops re-proposing them. Its baseline is the memory-based line Agents (LLM Agents) already tracks. No absolute figure, dataset or model is named in anything read, which is why it is here and not in Paper Picks (source)
• Trivium China's AI items were visible and unreadable for a second consecutive day. Three of today's 21 skipped Trivium candidates carry an AI subject — US accuses Chinese AI firms of industrial-scale distillation, China responds, the MIIT 15th Five-Year Plan for information and communications, and the SPC's judicial guidance on AI disputes — and triviumchina.com answers EGRESS_BLOCKED, so none was read past its title. A governance page is not written from a headline. This is a feed whose entries reach the candidate list and cannot reach a page (source)
• interests.md still has no row that reaches governance, and it decided today's ordering. Both Top Stories are scored on the default 1.0 topic weight because the table runs multimodal 1.0 → product launches 0.7 with nothing between for policy, regulation or procurement. It is the third consecutive day this gap has been recorded — after the VLOSE designation on 09-08 and the Trivium skips on 09-09 — and it is the reason a FOIA release of primary Pentagon documents scores 1.30, below five papers

For review: one person page and five paper pages.

• Paul Christiano (new — the wiki's first person holding both a lab safety-governance seat and a US model-evaluation post; review recommended)
• NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness (new)
• Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation (new)
• What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets (new)
• Steering Geometry: Validating Human Value Geometry in LLM Steering Space (new)
• Procedural Graphs: Self-Evolving Execution Structures for LLM Agents (new)

• OpenAI: two entries — the Foundation Board appointment, and the P00003 refusal-clause dispute. Also two Key People rows, for Christiano and for Zico Kolter, who chairs the committee
• Anthropic: the FOIA release's Anthropic items — the reported CENTCOM "target identification" use, and the refused classified follow-up. The supply-chain designation already on that page is referenced, not restated
• AI Governance: a new dated section for the disclosure and its shape — the first instrument on that page where the state and the lab are counterparties in both directions at once, every other one running regulator-to-provider
• Post-Training Scaling: two Key Papers entries, NeoHorse-1 and OPRD, with the teacher→verifier reading across three papers written into the second
• Agents (LLM Agents): two Key Papers / Events entries, for the production trading record and for Procedural Graphs
• Mechanistic Interpretability: a Key Papers entry for Steering Geometry
• index: six lines added — one person, five papers