ReFind is a keyword-based conversational search method achieving 58.2% accuracy without semantic indexing.

Ruizhe Li, Licheng Zhang, Benfeng Xu et al.

This paper identifies four structural barriers—web presence gap, token scarcity, tokenization penalty, connectivity exclusion—that hinder AI support for underrepresented languages like Bengali.

Avijit Roy, Proma Roy

This study reveals simulator collapse in multi-agent RL, proposing Verbalized Sampling and Co-Training to enhance generalization, improving success rates by up to 14%.

Simon Yu, Nicholas Tomlin, Marwa Abdulhai et al.

DARC uses failure diagnosis to restrict repair interventions, improving agent self-correction with 20-30% success rate gains.

Pan Wang, Yihao Hu, Hang Wang et al.

Proposes HPSE, a hybrid rollout self-distillation method, significantly improving unstructured knowledge editing's composability, with +6.8 average points across four LLMs.

Tianci Liu, Zihan Dong, Tianchun Li et al.

Proposes SpeedRunner, a programmatic skill learning method that reduces agent costs by analyzing trajectories without replay, achieving over 60% cost savings and performance gains.

Zixi Huang, Xiheng Wang, Andrew Wang et al.

This paper introduces a cross-lingual policy retention metric based on action traces, revealing that four models retain 71-73% of their strategies across languages after bias correction.

Sourabrata Mukherjee, Kalika Bali, Sunayana Sitaram

Introducing LingT2I benchmark, analyzing cross-lingual performance gaps and cultural influences in multilingual T2I models.

Sicheng Zhang, Zhonghao Yan, Binzhu Xie et al.

Combining Monte Carlo Dropout and input perturbations, the study assesses BERT variants' stability, finding RoBERTa most reliable with minimal output variance.

Pooja Yadav, Priyanka Harjule, Basant Agarwal et al.

Simplax enhances discrete diffusion models by introducing simplex-valued variables, improving generative performance.

Jinya Sakurai, Patrick Pynadath, Satoshi Hayakawa et al.

AlignXada uses verbal reinforcement learning for LLM personalization, achieving a 3.82-point gain while reducing 77.2% redundancy.

Yuting Liu, Wei Wu, Jianzhe Zhao et al.

ReMEMBER employs a gap-conditioned retrieval and evidence refinement framework, improving memory recall and gap resolution in long dialogue summarization with up to 160K tokens.

Hyangsuk Min, Hwanjun Song

Proposes Tied Trit-Planes, constraining PTQTP to a nine-level uniform quantizer with lossless folding, enabling efficient disk-streamed MoE serving with 4-bit codes.

Matteo Grella

Using Top-K sparse autoencoders, the study reveals that Thinking mode relies on sparse high-intensity features, while NoThinking favors symbol manipulation, with causal validation.

Bo Cheng, Qiaolin Lu, Yi Chang et al.

Introduces NCP-Bench, evaluating LLMs' long-horizon story consistency; GPT-5.2 achieves only 42% survival after 20 turns.

Yingpeng Ma, Jianhao Yan, Bei Shi et al.

PEP introduces trainable prompt embeddings into frozen LLMs for improved answer-level hallucination detection, outperforming standard linear probes in in-distribution settings.

Zakhar Mrykhin, Valentin Malykh

SurakshaEval evaluates multilingual LLM safety across 10 Indic languages, revealing cultural sensitivity issues.

Debopriyo Banerjee, Kapil Rajesh Kavitha, Angana Borah et al.

CoinRAG employs fine-grained nugget KV cache reuse to enable efficient long-context RAG, achieving a 5.3% relative F1 improvement under 100ms latency.

Gyuwan Kim, Cheoneum Park, Tao Yang

GPTKB 2.0 constructs a disambiguated knowledge base with 38.4 million triples, enabling traceable entity disambiguation using context-aware LLMs.

Yujia Hu, Tuan-Phong Nguyen, Simon Razniewski

Introduces the 'Horizon Gap,' analyzing LLMs' planning, memory, and execution challenges in long-horizon tasks via 1,547-paper survey.

Mingguang Chen, Licheng Wang, Bo Qu