Sequence-level distillation from Qwen3-14B raises small-model legal argument coverage, reaching ARCscore 0.781 on CANLII.

Mohamed Elaraby, Ahmed Elhady, Diane Litman

EVAR framework enhances evidence faithfulness and task performance in long-form narrative reasoning while controlling inference cost.

Peilin Liu, Zhiquan Ji, Jinglong Ping

ACTD employs vocabulary and sequence alignment with residual regularization, achieving state-of-the-art results on five reasoning benchmarks, with +4.5 points improvement.

Huiyi Zhang, Zijian Li, Xiaocheng Feng et al.

Detecting and guiding LLM-generated Korean poetry using interpretable features, achieving 83.60 AUC-ROC.

Keunhyeung Park, Seunguk Yu, YoungBin Kim

This work uses information theory to establish fundamental limits on extracting human intent solely from textual data.

Emily Cheng, Ryan Cotterell

This study compares sequential and parallel sampling in large language model translation, finding sequential sampling more effective under limited budgets.

Di Wu, Sergey Troshin, Christof Monz et al.

NL2AGBench evaluates large models’ ability to convert English geometry problems into AlphaGeometry formal language, achieving over 80% success in closed models.

Samuel Xiao, Judy Song, Rory Hu et al.

ElephantBench constructs a graph-based knowledge probe to diagnose LLMs' ability to recall multiple divergent accounts of long-tail facts, revealing biases towards dominant sources.

Zhuoshi Pan, Junru Lu, Yan Qian et al.

This study systematically analyzes how interlocutor biography visibility across training, inference, and evaluation affects persona-based dialogue generation, revealing leakage mechanisms.

Daniela Occhipinti, Malvina Nissim, Marco Guerini

This study introduces a diagnosis-guided post-training approach—Acquire, Repair, Preserve—that boosts small dialogue agents' scores from 10.67 to 38.92, focusing on local decision failures.

Nan Li

Sliding Window Attention (SWA) with sinks outperforms post-trained linear attention, enabling efficient long-context reasoning without extra training.

Alexia Jolicoeur-Martineau, Rhea Sanjay Sukthanker, Pashmina Cameron et al.

Proposes weak-model guided RLVR with external prefixes to enhance exploration, significantly expanding reasoning coverage and diversity.

Xingyu Shen, Huishuai Zhang, Peng Li et al.

This paper compares Merge, Mix RL, and MOPD in multi-domain RLVR, showing performance gaps up to 8.6 points and providing practical guidance for different scenarios.

Siye Wu, Kai Yang, Yuchen Cai et al.

Introduced DocTalkBN, a large-scale multimodal Bengali medical dialogue dataset with 557.63 hours, supporting triage, safety, and NER tasks.

Anik Saha, Fahmida Sultana Naznin, Sadatul Islam Sadi et al.

Cascaded Batch Prompting improves performance by disentangling complex reasoning from symbol grounding.

Sho Hoshino, Peinan Zhang

KinyaEmbed improves Kinyarwanda sentence embeddings via a four-stage curriculum, achieving Spearman ρ=0.7298.

Ireddi Rakshitha, Devavarapu Yashwanth, Ntakirutimana Pierre

MoganColBERT-TR uses MaxSim for token-level retrieval, achieving strong zero-shot performance on five Turkish BEIR datasets, outperforming larger models.

Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozay

DEDUCE framework enhances LLM robustness against input fact perturbations, showing a 20% accuracy improvement in experiments.

Ping Wang, Xiangguo Sun, Bingbing Xu et al.

TOPAS jointly optimizes prefix residency and request scheduling, reducing task JCT by up to 39.8% on synthetic workloads.

Hongqiu Ni, Han Tian, Chi Zhang et al.

Unified Bayesian model for RAG evaluation, decomposing retrieval success, abstention, and correctness; reveals behavioral differences across 27 configurations.

Pius von Däniken, Felix Matthias Saaro, Mark Cieliebak et al.