$ cat briefs/daily/2026-11-14.md
2026-11-14
November 14, 2026 (Sat)
1 top story · 1 paper pick · 1 evidence-auditing signal
**Reconstructed archive:** prepared on **2026-11-15 (KST)** after the original automation failed, using the original morning cutoff of **2026-11-14T08:00:00+09:00** (**2026-11-13 23:00 UTC**). This issue was not published or emailed on the missed date; no subscriber delivery is authorized. **Provenance and limits:** the research below is preserved in the **2026-11-13** Hugging Face snapshot. Its bytes match repository commit `ed9d75b58e38782c30a0ee4ab111b6429e630652`, timestamped **2026-11-13 00:25:08 UTC**, before the cutoff. Version-specific primary records below confirm earlier submissions; these are not releases on the issue date. Coverage is limited to abstracts and metadata; full papers, code and results were not independently validated. This selective reconstruction cannot recover the complete morning feed. Items without verified pre-cutoff availability are omitted. Wiki links provide retrospective topic context and may contain later updates. ([snapshot](../../sources/papers-daily/hf-daily-2026-11-13.md))
Top Stories
1. SteerDistill trains quantized agents to recover from behavioral errors
A full-precision teacher helps a quantized student revise failed responses, refining its guidance using feedback from execution. The student then learns from the corrected behavior for inference without guidance. Across 5 agentic benchmarks and 3 models, authors report 1.67–2.18 points higher average scores than the strongest quantized baseline under W4A16, staying within 1 point of BF16.
Why it matters: corrections drawn from the student's own failures can address contexts missed by offline training data. These are author-reported benchmark comparisons, not independently verified deployment gains. Version history: v1, 2026-11-10 00:45:58 UTC. → Agentic Reinforcement Learning (snapshot) (primary v1)
Paper Picks
Tree-Ising: repair reasoning loops through sparse bit-level weight edits, optimizing an objective derived from forward-pass statistics. On Qwen3.5-0.8B, authors report 54.6% on MATH-500 versus 51.7% for an agent-produced edit under the same evaluation protocol; the reported base-model score is 7.9%. Why read it: structured optimization offers a comparison point for agent-written weight repairs. This narrow experiment does not establish a general replacement for model training. Version history: v1, 2026-11-10 15:32:04 UTC. → Agents (LLM Agents) (snapshot) (primary v1)
Watch
- BioAgent-DF: controlled changes to saved computation values altered the evidence agents read or cited while final verdicts remained unchanged. Authors report 7 repeatable capture-claim combinations covering 2 claims across their SciFact experiments. Why it matters: final-answer agreement can hide changes in the supporting evidence; these controlled results motivate provenance audits without establishing their prevalence in deployed systems. Version history: v1, 2026-11-01 03:58:26 UTC. → Eval Harness Configuration (snapshot) (primary v1)
New in Wiki
- Original dated work: the retained log records creation of the Harness-S and Hybrid Computer Use paper pages on 2026-11-08, not 2026-11-14. Reconstruction: no wiki pages were created; this issue links existing topic pages. (log)
Updates
- Original dated work: the 2026-11-08 ingest entry records updates to Agents (LLM Agents) and Agentic Reinforcement Learning; no 2026-11-14 ingest entry is retained. Reconstruction: no wiki pages were edited. An absent entry does not establish an absence of news. (log)