The results show that explicitly selecting and integrating evidence improves long-context reasoning while enabling more compact generation, and the strongest overall result among the evaluated open-source models is obtained.
Abstract
Long-context understanding requires large language models (LLMs) to reason over lengthy documents, conversations, and code, yet task-relevant evidence is often sparse and scattered amid substantial irrelevant and redundant content. We propose Highlight-Then-Summarize (H2S), a compress-then-reason paradigm that first identifies source-grounded, question-relevant evidence and then integrates it into a compact, question-conditioned summary before producing the final answer. To train this behavior, we construct H2S-Dataset, comprising 6,647 examples from 11 benchmark families with an average context length of 43.9K tokens, and introduce H2S-RL, which provides process-level rewards for evidence selection and summary construction in addition to final-answer correctness. We evaluate on H2S-Bench, a seven-task long-context suite. Under a shared 128K input and 4K output budget, H2S-14B achieves an average score of 32.60, outperforming Qwen3.8-27B by 10.17 points and obtaining the strongest overall result among the evaluated open-source models. H2S-14B also achieves the highest Evidence-Summary Quality score and retains 97.1% of its 16K-budget performance with only a 4K output budget. These results show that explicitly selecting and integrating evidence improves long-context reasoning while enabling more compact generation.
SCSP is a training-free framework that leverages semantic compression for informative token selection in long-context embedding and can be integrated into both zero-shot and fine-tuned models in a plug-and-play manner, consistently improving their performance.
Zi-Feng Cheng, Jie Zheng, Zhiwei Jiang et al.· 0 citations
This work introduces DocHop, a benchmark for integrated chart--context reasoning in document-style images and constructs DocHop via a stochastic logic-first generation pipeline with controllable reasoning depth and visual density, to enable systematic evaluation.
Zhuo-Ran Yu, Le Thien Phuc Nguyen, Jaden Park et al.· 1 citation
This work proposes MemSuit, a self-distillation framework in which a teacher summarizer, conditioned on observed query-answer pairs, produces utility-aware memory entries that a student learns to reproduce from the raw conversation alone, confirming the value of grounding memory in downstream utility.
Zhen-Yu Lei, Ming-Jia Shi, Xing-Bo Fu et al.· Proceedings of the 2026 Conf...· 0 citations
This work attempts to make use of high-quality public documents with small perturbations and empirically finds that LLMs can successfully generate context-dependent reasoning traces and answers, which are then used to train a student model that substantially improves the performance on CL-bench.
Hao Wu, Yang Xiao, Yu-Song Sun et al.· 0 citations
This paper describes the system for Task~2 of the second Multilingual Conversational Speech Language Model (MLC-SLM) Challenge, which adapt Qwen3-Omni-30B-A3B-Instruct with a segment-evidence-aware data and post-training pipeline and obtains 90.92% accuracy on the final official evaluation set.
H. Le, L. Nguyen, Minh Tri Dao· 2nd Workshop on Multilingual...· 1 citation
CoEM preserves potentially useful source excerpts verbatim in a pending set, allowing subsequent context to clarify their relevance before irreversible compression, and consistently improves long-context reasoning.