Tag
This ICAIF '26 paper studies hindsight self-distillation for multi-step financial reasoning on GPT-OSS-20B, finding that a fixed generic feedback line matches the ~3-point gain of self-written critiques, suggesting the positional information in hindsight distillation reduces the need for privileged feedback.
该论文提出 Fin-SD（hindsight self-distillation），一种无需外部教师模型的本地化自蒸馏方法，用于提升小型开放权重模型的多步金融推理能力；实验发现增益主要来自监督位置而非提示内容，且与基于完整参考解答的自蒸馏精度相当但少生成 9.5% token。
提出Hindsight Structured Advice Distillation（HSAD），用一个小型advisor模型根据输入生成结构化建议并注入冻结的executor agent提示中，advisor通过分析执行失败与任务特征进行监督微调训练；在FrontierFinance基准上将Samaya研究agent的合格率从52.9%提升至56.6%，且成本更低，并在WebShop、BIRD和ExpertLongBench上验证了泛化性。
This paper introduces a data-centric pipeline for post-training language models to enhance financial reasoning through mining reasoning traces, distilling instruction data, and generating verifiable QA pairs, demonstrating improvements in performance while preventing catastrophic forgetting.
This paper presents HSDN, a hierarchical semantic distillation network for compressing chain-of-thought reasoning in financial tasks, achieving 91.0% accuracy with 68.4% compression on the AFAC2025 benchmark.
Introduces FinIndices, a large-scale benchmark evaluating LLM data-processing fidelity on uncropped financial statements, revealing knowledge and structural bottlenecks in financial reasoning.
MacroLens is a new multi-task benchmark for contextual financial reasoning that jointly evaluates price history, accounting fundamentals, macroeconomic regimes, and textual data across 4,416 U.S. small- and micro-cap equities. It includes seven tasks, 1,130 macroeconomic events, and evaluations of 19 methods, aiming to fill a gap in financial AI evaluation.
FinAcumen is a framework that accumulates reasoning experience from prior trajectories into a persistent memory bank for financial multimodal reasoning, improving performance across four benchmarks while maintaining a frozen 8B vision-language model.