Negative Self-Distillation: Learning to Reason by Avoiding Flaws

简体中文

Local core-mechanism evidence. Paper outcomes and local results are separate.

Paper information

Field	Value
Paper	Negative Self-Distillation: Learning to Reason by Avoiding Flaws
Affiliations	University of Virginia
First public release	2026-09-10
Authors' code	Authors' code.
Adapter / method	nsd
Local code	src/auto_research/post_training/latest_20260914.py

Original paper

Background and contribution

Move the student away from its own flawed reasoning distribution. A dynamic gate limits updates to reasoning-relevant positions to avoid forgetting ordinary language.

flowchart LR
  I[Candidate policy] --> N[Negative teacher from flaws] --> G[Dynamic reasoning gate] --> U[Selective update]

Key figure from the paper

Original Figure 1: Negative Self-Distillation. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

Construct the negative teacher and dynamic reasoning-position gate. Tests inspect gate, selection, and state invariants.

Paper results

Average gains on seven math benchmarks for 1.7B/4B/8B were 2.3%/7.5%/6.0% in the paper. No local comparison or inferred online A/B.

Local reproduction

Three-seed metrics use matched public fixtures, budgets, and seeds. The deterministic kernel exposes gates and update statistics.

Reproduction limits

L1 negative-teacher/gate diagnostic only; no full-parameter LLM training.