AdviSD: Learning to Advise Frontier LLMs via Targeted Multi-Turn Self-Distillation¶
L1 paired contrast, donor calibration, and selection gate; no Gemini/Claude executor or BFCL-v3 run.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2609.38142 |
| Institution | University of Southern California (first author). |
| First public release | 2026-09-29 (arXiv v1) |
| Author code | No author repository found as of 2026-10-01. |
| Local adapter or method key | advisd |
| Local reproduction code | src/auto_research/post_training/latest_20261001.py |
Paper method and evidence¶
Freeze the executor and train the advisor. Reflection proposes a correction; score the same logged executor response with and without advice. Distill only decisions with sufficient effect under donor-advice calibration.
flowchart LR
R[Reflect and propose advice] --> C[Score same response with and without advice]
C --> G[Donor-calibrated effect gate]
G --> T[Distill selected decisions]Original paper figure¶
Figure 1 shows the distillation workflow. From the original paper; copyright remains with the authors.
Core equation¶
\(c_k=T_k^{-1}\sum_t[\log\pi(y_{k,t}|C_k^+)-\log\pi(y_{k,t}|C_k^-)]\). Threshold on an empirical quantile of absolute donor contrasts.
Reported results¶
The paper reports 4.2–6.4 percentage points on BFCL-v3 and 3.9–5.1 points on EnvScaler versus advisor-GRPO.
Local reproduction¶
Compare with the corresponding operator disabled or in its default state. Three-seed metrics are diagnostic_only=true: they verify execution and invariants, not formal gains or capability ranking.