Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning

简体中文

L1 neighborhood perturbation and stability penalty; no actual frozen LLM prefix loaded.

Paper information

Field	Details
Paper	Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning
Institution	Chongqing University (first author).
First public release	2026-09-10 (arXiv v1)
Author code	No public author repository found as of 2026-09-14.
Local adapter or method key	tasco
Local reproduction code	src/auto_research/post_training/latest_20260914.py

Paper method and evidence

Freeze the base model and optimize a lightweight prefix. Penalize instability under neighborhood perturbations as well as low confidence, to avoid confident but incorrect trajectories.

flowchart LR
  P[Candidate prefix] --> N[Random neighborhood perturbations]
  N --> S[Confidence and stability objective]
  S --> U[Prefix-only update]

Original paper figure

Figure 3 shows the test-time adaptation process. From the original paper; copyright remains with the authors.

Core equation

The deterministic reference emits decisions and state statistics and tests mathematical invariants and boundaries.

Reported results

Paper results are author measurements, not fixture scores.

Local reproduction

Three-seed mechanism results are diagnostic_only=true. Fixed-budget fixtures verify the defining operator and auditable intermediate states; they are not a benchmark, production result, or capability-ranking entry. All candidates use the same public fixtures, budget, and seeds.