arxiv arXiv cs.CL · just now · src: 1d ago · research

R3Con applies relevance realization principles to improve large-scale reasoning

from English

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.

Evaluated against nine baselines on two benchmarks for reasoning over large document corpora.
Outperforms the strongest baseline by 20 and 8.4 percentage points respectively.
Enables smaller models to outperform larger ones: R3Con with 4B and 9B models beats all evaluated 35B baselines.
R3Con with a 35B-A3B model outperforms Claude Code with Claude-Sonnet-5 at 3.7x lower cost.

The results indicate that context representations following these principled approaches can reduce reliance on model scale, pointing toward frontier-level performance powered by smaller models.

Importance 2/3 New harness with differentiators arXiv cs.CL Research paper Retrieval & RAG

Read original

media Hugging Face Forums · 27d ago · 38 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.

arxiv arXiv cs.CL · 70d ago · 22 views

Apaf hybrid framework analyzes discourse in disaster governance policies

The authors present Apaf, a hybrid LLM-symbolic pipeline that operationalizes critical discourse analysis as a quantitative bipolar argumentation framework over policy text. Arguments are classified into deliberative or managerial frames, and four frame-mediated relation subtypes are produced by deterministic rules over LLM-extracted features.

arxiv arXiv cs.CL · 70d ago · 24 views

Theory-Level Autoformalization: From Isolated Statements to Unified Formal Knowledge Bases

This position paper argues for theory-level autoformalization, which involves formalizing complete theories as structured libraries rather than focusing on individual statements.

arxiv arXiv cs.CL · 70d ago · 17 views

Corpus-adaptive fine-tuning improves SQL schema retrieval recall

The authors recast five text-to-SQL datasets (Spider, BIRD, BEAVER, and two LiveSQLBench variants) as table and column retrieval tasks to evaluate schema linking. They propose corpus-adaptive fine-tuning for a 305M-parameter embedder, which raises average recall@10 from 60.4 to 75.6.

arxiv arXiv cs.CL · 70d ago · 24 views

POS tier halves data needed for automated interlinear glossing of Irabu Ryukyuan

Researchers implemented a neural annotation pipeline for Irabu Ryukyuan to automate the creation of interlinearized text, addressing the high cost of manual documentation. The study evaluated morpheme segmentation, POS tagging, and glossing using small BiLSTM-CRF models under a strict constraint of approximately one hour of annotated discourse.