HISD：用提示自蒸馏改进推理分割的 token 级信用分配
Hint-Informed Self-Distillation for Reasoning Segmentation
做视觉推理或分割模型 RL 训练的可以看看这篇：HISD 用 EMA 教师在 token 级分配证据，比 DPAD 在 ReasonSeg 上 gIoU 高 3 个点，推理 token 还砍掉八成以上，而且部署时不带教师，成本不涨。
新加坡国立大学等团队在 arXiv 论文（编号 2611.14931）中提出 Hint-Informed Self-Distillation（HISD），解决推理分割中轨迹级奖励无法定位哪些 token 提供关键视觉证据的问题。HISD 先由外部模型离线生成 Segment Hints，内容涵盖目标身份、判别性属性、易混淆物体和锚定路径；再由提示条件下的指数移动平均（EMA）教师对学生自身轨迹打分，通过 token 支持集上的师生证据比率调整正组相对优势的幅度，非正优势保持不变。
在统一训练协议下，HISD 相对作者复现的 DPAD 基线，在 ReasonSeg 测试集上将 gIoU 和 cIoU 分别提升 3.00 和 3.55 个百分点；在 ReasonSeg 五个评测划分和 RefCOCO 系列上均超过该基线，同时把有效推理跨度的平均 token 数削减 80.17% 到 91.83%。推理阶段学生模型既不需要提示也不需要教师模型，部署成本不变。
这项工作把奖励建模从轨迹级细化到 token 级，对从事视觉推理和分割模型 RL 训练的开发者有直接参考价值，短推理、高精度的组合也可能降低下游部署的推理开销。
Hint-Informed Self-Distillation for Reasoning Segmentation
Reasoning segmentation requires identifying the target referred to by a query among visually similar objects, yet trajectory-level rewards cannot directly indicate which tokens provide relevant visual evidence. We propose Hint-Informed Self-Distillation (HISD), which uses target-disambiguation evidence for token-level credit assignment. An external model generates Segment Hints offline from training images, queries, and ground-truth target annotations, organizing target identity, discriminative attributes, confusable objects, and anchoring paths. A hint-conditioned exponential moving average teacher scores the student's own trajectories, and HISD computes the teacher-to-student evidence ratio over a token support set to adjust only the magnitudes of positive group-relative advantages, leaving nonpositive advantages unchanged. This design encourages concise reasoning while retaining the information needed for target disambiguation, and the student requires neither hints nor a teacher at inference. Under a unified training protocol, HISD improves ReasonSeg test gIoU and cIoU over our reproduction of DPAD by 3.00 and 3.55 percentage points, respectively. Across the five evaluated splits of ReasonSeg and the RefCOCO family, HISD achieves higher segmentation metrics than this baseline and reduces the average token count of valid reasoning spans by 80.17 to 91.83 percent.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include HISD, ReasonSeg, RefCOCO, 推理分割.
Source attribution: curated from arXiv cs.AI (@Kunze Jiang, Yuchen Wang, Kang Li, Qixing Zhang).