Not All Tokens Deserve Equal Credit: Counterfactual Sensitivity Credit Reallocation for Long-CoT Reasoning

  • 类型：arxiv
  • 标识：2607.27888
  • 链接：https://arxiv.org/abs/2607.27888
  • 主分类：multimodal
  • 形态：method
  • 被引：0
  • 被引来源：Semantic Scholar + OpenAlex
  • S2被引：0
  • OpenAlex被引：0
  • 影响力被引：0
  • TLDR：These findings show that privileged shifts fail to provide reliable answer-aligned directions, while their magnitudes primarily reflect counterfactual sensitivity rather than token-level learning value, and propose Counterfactual Sensitivity Credit Reallocation (CSCR), a simple extension of GRPO that reduces credit for highly sensitive tokens and renormalizes token-level advantages to preserve both the original credit budget and verifier-determined direction.
  • OpenAlex ID：W7171973840
  • OpenAlex DOI：10.48550/arxiv.2607.27888
  • DOI：10.48550/arxiv.2607.27888
  • DOI来源：OpenAlex
  • 开放获取：green
  • 开放获取链接：https://doi.org/10.48550/arxiv.2607.27888
  • OpenAlex更新：2026-08-26
  • 待LLM分类：否
  • 标题中文：并非所有 token 都应获得同等贡献：面向长链思维推理的反事实敏感度贡献重分配
  • TLDR中文：研究发现privileged shifts无法给出可靠的答案对齐方向，其幅度主要反映反事实敏感性而非token级学习价值；提出Counterfactual Sensitivity Credit Reallocation (CSCR)，作为GRPO的简单扩展，降低高敏感token的credit并对token级优势重新归一化，同时保留原始credit预算与verifier确定的方向。
  • 来源文件：
  • /inbox/tom/_candidates/2026-08-03-agent-rag-longcontext-candidates.json
  • /inbox/tom/_candidates/2026-08-04-rag-retrieval-reranking-candidates.json
  • /inbox/tom/_candidates/2026-08-04-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]