This paper systematically analyzes lossy verification in speculative decoding, classifying into truncation and collaborative methods, revealing their mechanisms, pitfalls, and control principles.

Tianyu Wang, Yuxuan Zhou, Wenbin Wang et al.

Proposes a human-in-the-loop workflow using SciSummNet and GPT-4o-mini for scientific summary simplification, with expert editing to preserve terminology and claims.

Kyuri Im, Michael Färber

Proposes an instruction-free alignment approach for audio-language models, training only a lightweight projector, achieving performance comparable to heavily fine-tuned models.

Xuanru Zhou, Yiwen Shao, Jiahong Li et al.

GEMCo uses human-crafted proxy conversations validated against real data, enabling ethical sharing of sensitive counseling dialogues with minimal distribution gap.

Philipp Steigerwald, Eric Rudolph, Mara Stieler et al.

Study finds Chinese debates show higher semantic repetition than English, introducing 'Prior-Argument Similarity' metric.

Huiqian Lai

IKS-Instruct creates a 24,795-pair multilingual dataset covering 41 Indian Knowledge System techniques, significantly enhancing model cultural depth and pedagogical ability.

Shwetha Singaravelu, Gayathri Muruganantham, Lakshmi Rajendran et al.

Proposes BHARATI, a morphology-aware tokenizer with subword fertility analysis, reducing sequence length by 90% on classical Indian language corpora.

Poornima Kumaresan, Pavithra Muruganantham, Lakshmi Rajendran et al.

Proposes Co-E system with bidirectional graph-text memory for training-free multi-hop QA, outperforming baselines.

Hieu Man, Thien Huu Nguyen

This paper derives power-law scaling laws for from-scratch vision-language pretraining, revealing distinct behaviors for language and multimodal objectives under fixed compute.

Haoyuan Wu, Aoqi Wu, Hai Wang et al.

Unified moral-value dataset for instruction tuning enhances AI's ethical alignment, improving value-behavior consistency.

Zhaohui Zeng, Florian Mai

REFACT enhances long-text reasoning by adaptive fact restatement, reducing tokens and improving evidence density.

Zhensheng Jin, Xin Dai, Zhenghao Liu et al.

Proposes two reliability-aware distillation methods, CHAD and EWAD+CPDP, outperforming standard KD with +0.0219 ROUGE-L on low-resource summarization.

Dipto Sumit, Ankan Kumar Roy Srizon, Sadia Khair Rodela et al.

Proposes Sentence Splitter, a T5-based model that automatically identifies factual sentence boundaries, improving knowledge graph completion and QA by 4-6%.

Ahmad Pouramini, Mahsa Afsharizadeh

Multilingual sentence embeddings replace translation in LiRA, recovering responses excluded by translation failures with <1% reliability difference.

Ummugul Bezirhan, Ji Yoon Jung, Matthias von Davier

This study distinguishes OC and SJ signals, showing SJ directions transfer better across domains, highlighting semantic ambiguity in correctness probes.

Yi-Long Lu

Using mass-mean probing, the study demonstrates that neural language models encode grammaticality as a linearly separable dimension in sentence representations, with success rates from 13.4% to 94.0%.

Jane Li, Najoung Kim

D-cut improves high-concurrency inference speed by adaptive pruning, achieving up to 3× acceleration.

Tianyu Liu, Yuhao Shen, Rui Cen et al.

Trained small language models using progressive SFT and RL to enhance multi-agent retrieval relevance, achieving NDCG@10 of 0.918.

Gayathri V Kondapalli, Alexander Ng, Hirsh Pithadia et al.

HarnessBank achieves agent-harness self-evolution with semantic gene-bank search and gated verification, improving performance by 5.1%-15.4%.

Xiaotian Luo, Dizhan Xue, Fengxingyu Wang et al.

MemCon models memory control as an MDP, using lightweight online UCB policy, achieving 15.2% success boost and 20% token savings without extra LLM calls.

Eric Hanchen Jiang, Zhi Zhang, Yuchen Wu et al.