arxiv arXiv cs.AI · 3d ago · src: 5d ago · research

Highlight-Then-Summarize compresses evidence to improve long-context understanding

from English

Researchers propose Highlight-Then-Summarize (H2S), a compress-then-reason paradigm that identifies question-relevant evidence and integrates it into a compact summary before generating an answer. The team constructs H2S-Dataset with 6,647 examples and introduces H2S-RL to provide process-level rewards for evidence selection.

H2S-14B achieves an average score of 32.60 on the seven-task H2S-Bench suite under a shared 128K input and 4K output budget.
The model outperforms Qwen3.8-27B by 10.17 points, obtaining the strongest overall result among evaluated open-source models.
H2S-14B retains 97.1% of its 16K-budget performance with only a 4K output budget while achieving the highest Evidence-Summary Quality score.

These results demonstrate that explicitly selecting and integrating evidence improves long-context reasoning while enabling more compact generation.

Importance 2/3 arXiv cs.AI Research paper Retrieval & RAG

Read original

arxiv arXiv cs.AI · 2d ago · 1 view

RareDx uses graph-grounded policy optimization to improve rare-disease diagnosis

The authors introduce RareDx, a system that couples controlled evidence use with knowledge-graph-grounded policy optimization to address the long-tail reasoning problem of rare-disease diagnosis. The training pipeline combines Top-10 post-training with RareDx-KGPO, which projects predictions into a canonical disease graph and integrates curated graded relevance, ontology proximity, biomedical similarity, and phenotype consistency.

arxiv arXiv cs.CL · 3d ago · 6 views

Highlight-Then-Summarize compresses evidence for better long-context understanding

Researchers propose Highlight-Then-Summarize (H2S), a compress-then-reason paradigm that identifies question-relevant evidence and integrates it into a compact summary before generating an answer. To support this approach, the team constructed H2S-Dataset with 6,647 examples from 11 benchmark families and introduced H2S-RL for process-level rewards.

media Hugging Face Forums · 3d ago · 5 views

Biological JEPA models disease progression using biological constraints

A new open-source project called Biological JEJA combines JEPA-style representation learning with biological constraints to model how diseases progress over time. The approach was tested on real ADNI data from 2,347 patients and compared against several other models.

arxiv arXiv cs.CL · 7d ago · 6 views

R3Con applies relevance realization principles to improve large-scale reasoning

Researchers propose R3Con, a harness that operationalizes cognitive theory principles for constructing effective representations of very large contexts. The system addresses the challenge of assembling interdependent information scattered across vast sources beyond standard model context limits.

media Hugging Face Forums · 34d ago · 44 views

Ontological inversion: cross-model vector memory for steering and recall

A preprint by independent researcher Jason Van Pham proposes translating continuous representations between frozen models to enable steering, recall, and reasoning without weight editing. The work introduces a method where concepts are compiled into residual directions in one model's space and mapped via an affine transformation into another model's hidden layers.