[2026-10-04] Self-Evaluating Recursive Agents
SERA 用强化学习训练同一个 Qwen3-4B 模型拆任务、执行和评价。关键对照是：只加评价训练不够，还要把学到的 rubric 用作子任务执行奖励。本文分开解释训练、普通评测和测试时选择，并对照代码说明运行框架…
淡水鱼的蔚蓝之海
课程笔记、踩坑记录，还有一些摸鱼和日常。
复仇者不折镆干。虽有忮心，不怨飘瓦。
[2026-01-26] Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models
OPSD 让同一个模型在两种上下文下分别充当学生和 teacher：学生只看问题并生成自己的轨迹，teacher 额外看到 verified solution，再给同一批 token 密集监督。本文同时核对论文与官…
[2026-02-12] On-Policy Context Distillation for Language Models
OPCD 研究怎样把历史经验或优化过的 system prompt 蒸馏进模型参数。它保留 student 的 on-policy rollout，再让 context-conditioned teacher 对同…
[2026-02-04] Privileged Information Distillation for Language Models
π-Distill 面向多轮 agent：frontier teacher 不开放 hidden reasoning，只留下成功的 action/tool-call trace。论文研究怎样把这类 action-o…
[2026-09-22] What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillation
这篇系统比较 self-teacher 的五档 privileged context，从 full solution 到 answer-only。本文逐项核对 29,434 行公开数据、L2-L4 compiler…
[2026-05-09] Training with Harnesses: On-Policy Harness Self-Distillation for Complex Reasoning
OPHSD 把临时 inference harness 变成训练期 teacher context：student 仍直接作答，teacher 则读取 Plan–Solve 或 Draft–Verify 的 ter…
[2026-06-09] Self-Distillation Policy Optimization via Visual Feedback: Bridging Code and Visual Artifacts
Visual-SDPO 执行 student 代码，把 screenshot、rubric 与 runtime error 作为 teacher-only context，再把视觉 defect 定位回负责的 co…
[2026-06-04] ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation
ViCuR 区分 answer-side privilege 与可从原图恢复的 visual cue，并用 SinkTrack 让一个 sink token 定期读取全体 visual tokens。它适合帮助定义…
[2026-08-05] Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation
这篇反例论文说明 PI-conditioned teacher 可能奖励 reference-path similarity，而不是 correctness。它对 QA、数学、代码和多轮 tool use 的 de…
[2026-08-03] Is More Privileged Information Better? From Solution Traces to Problem-Solving Structure in Self-Distilled Reasoning
PS-OPSD 不把完整 solution 交给 self-teacher，而是提取 initial state、goal、constraints 与 state-transition path。它说明结构化 pr…
