SP-DocReader introduces a self‑play framework for OCR that focuses on residual errors after supervised fine‑tuning. It uses Reading Discrepancy Masking to align reference and generated tokens via a longest common subsequence and scores unmatched positions with full conditioning prefixes. The method adds a Focused Fidelity Loss for direct negative log‑likelihood supervision at unmatched ground‑truth positions, training only the OCR module while keeping the backbone frozen, and achieves significant character error rate reductions and ANLS improvements on multiple backbones.
By Wenjie Liao, Xiaohui Song, Liangjie Zhao, Haonan Lu
The paper introduces Harness Compilation (HC), an offline method that reallocates tasks between a frozen small vision‑language model (VLM) and an external harness without updating model weights or requiring teacher calls. HC uses a large teacher to refine reusable content and control based on student execution traces, then selects the best harness from a validation set. Across seven visual question‑answering benchmarks, HC boosts performance of VLMs up to 23.9 points and demonstrates that selective delegation of decisions—such as evidence requests and query generation—can be more efficient than uniform removal of all decisions.
By Minhao Fan, Yinyi Liu, Jiayu Zhao, Zihan Teng, Song Chen, Weichen Liu
The paper introduces Multi-Agent Self-Supervision (MASS), a recursive self‑improvement technique that alternates between evolutionary workflow optimization and supervised fine‑tuning on self‑generated multi‑agent trajectories. MASS enables a single base model to iteratively propose, execute, and self‑evaluate complex multi‑agent workflows, discovering effective roles and information routing. Experiments with Qwen3.6‑27B show 1.2–1.6× higher performance per output token on four open‑ended benchmarks, and student models trained on MASS traces outperform single‑agent students trained on 1.4× more tokens.
By Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia, Yujin Tang
The study examines whether large language models truly base legal verdicts on the statutes or precedents they cite. By fixing case facts and swapping the cited authority with an unrelated one, researchers found that while models correctly name the authority 66.7%-100% of the time, the resulting verdict often remains unchanged—only 0.0%-21.7% of the time on CaseHOLD, 30.0%-76.7% on ECHR and SCOTUS, and 43.3%-50.0% on ContractNLI. Even a purpose‑built legal‑reasoning model and larger scales did not close this gap, and adversarial instructions further exposed vulnerabilities in verdict consistency.
By Saisab Sadhu, Shreeyans Arora, Pratinav Seth
Nullify is a training‑free method for unlearning sensitive information from large language models by steering activations during inference. It uses steering vectors to redirect privacy‑related activations away from memorized answers while enforcing a null‑space constraint that preserves activations for retained queries, thereby maintaining model utility. Experiments on TOFU and MUSE demonstrate that Nullify matches or outperforms existing baselines in forget quality while keeping utility nearly intact, all without any weight updates.
By Wei Zhai, Xiang Liu, Qiang Huang, Rui Qian, Lemao Liu, Ziwei Li, Ziqi Wang, Zhitao Huang, Dejing Dou
The paper introduces SFT-as-context, a training‑free method that lets a parent large language model (LLM) use the response of a supervised fine‑tuned (SFT) model as context to answer queries. This approach preserves the parent model’s general capabilities while acquiring the SFT model’s specialized skills, achieving performance close to the SFT model on fine‑tuned benchmarks and near the parent model on general tasks. Experiments across 19 parent‑SFT pairs and 11 benchmarks show minimal gaps in performance, and the method can solve problems requiring both specialized and general knowledge, even when neither model alone succeeds.
By Kenan Tang, Andong Hua, Chengxuan Qian, Saket Tiwari, Yao Qin
MAP4CS is a data pruning framework that selects a small, high‑quality core subset from large code corpora by combining syntactic structure, semantic diversity, and distributional representation, followed by rule‑based filtering. Experiments on two large datasets show that using only 5% of the training data, MAP4CS outperforms random sampling and matches or exceeds full‑corpus fine‑tuning, supporting the idea that less data can be more effective. The method also automatically removes duplicates and noise, producing a lexically diverse and information‑dense training set.
By Yuxuan Chen, Mingwei Liu, Guangsheng Ou, Zekai Zhang, Zike Li, Yanlin Wang, Pelin Zheng
Syn-Omni introduces a structured approach to omnimodal embeddings by separating universal semantics from modality-specific features through Orthogonal Modality-Expert LoRA (OME-LoRA). It further employs Progressive Synergy Routing (PSR) to allow modality experts to first learn modality-specific priors and then gradually collaborate across modalities. Across 81 tasks involving image, video, audio, and audiovisual data, Syn-Omni outperforms existing omnimodal baselines, highlighting the benefits of structured specialization and progressive cross-modal collaboration.
By Youngtaek Oh, Qiyu Wu, Hiromi Wakaki, Junmo Kim, Yuki Mitsufuji
The paper introduces Trust Region Q Adjoint Matching (TRQAM), an off‑policy fine‑tuning algorithm that stabilizes reinforcement learning of pretrained flow policies by adaptively controlling the path‑space KL divergence between the fine‑tuned and pretrained policies. TRQAM achieves this through projected dual descent, which adjusts a trust‑region parameter λ during sampling and ensures the KL penalty is exactly weighted in the stochastic optimal control objective. Experiments on 50 OGBench tasks show TRQAM consistently outperforms prior methods, achieving a 68% success rate in offline RL versus 46% for the strongest baseline.
By Yonghoon Dong, Kyungmin Lee, Changyeon Kim, Jaehyuk Kim, Jinwoo Shin
The paper introduces MGRASRec, a multimodal graph retrieval-augmented framework that enhances sequential recommendation by incorporating collaborative filtering signals directly into a multimodal large language model (MLLM) prompt. It retrieves structured paths from a user-item interaction graph, extending them with multimodal similarity to broaden coverage and surface the most relevant history items for each candidate item. This approach eliminates the need for recurrent summarization, requiring only a single forward pass per candidate and achieving superior performance across three datasets, especially in ranking quality.
By Jason Marcell Setiadi, Xin Cao, Lina Yao
FloorSAV is a framework that enhances audio‑visual large language models (AV‑LLMs) by embedding a dynamic 2D floormap derived from 3D point clouds, camera trajectories, spatial audio cues, and object landmarks. The floormap is fed into the AV‑LLM as a synchronized stream with egocentric video, enabling the model to jointly reason over visual, auditory, and geometric information in a single inference. The authors also introduce SAVED‑Bench, a benchmark for spatial reasoning tasks, and show that FloorSAV improves AV‑LLM performance on both SAVED‑Bench and SAVVY‑Bench, with ground‑truth floormaps highlighting its potential.
By Kyeong-Rae Kim, Sungnyun Kim, Tae-Hyun Oh
CMGL is a confidence‑guided multi‑omics graph learning framework that first estimates modality confidence for each patient using evidential deep learning, then fuses omics data with fixed confidence values and performs classification on a consistency graph. On four cancer‑subtype tasks and a 32‑class pan‑cancer task, CMGL outperforms the strongest baseline by an average of 4.03% accuracy. Its learned representations recover known breast cancer subtypes and transfer to kidney cancer, enabling prognostically distinct patient stratification.
By Boyang Fan, Hengchuang Yin, Siyu Yi, Yifan Wang, Zhicheng Li, Leijiyu Zhou, Jiancheng Lv, Wei Ju
The paper investigates how byte‑level language models that use entropy‑based patching (like BLT) fail to compute certain predictable but value‑dependent positions, such as numbers following an equals sign in math solutions. By introducing a boundary‑dependence signal—measuring the model’s loss when a patch start is removed—they show that this signal, combined with entropy, significantly improves accuracy on computed outputs across various model sizes and budgets. The study demonstrates that entropy alone leaves a systematic blind spot that boundary dependence can fill, leading to higher exact‑match rates on challenging computed tasks.
By Nicol\'as Vera Z\'u\~niga
EgoVoice is a framework designed to train and evaluate proactive spoken assistants that operate from egocentric multimodal streams. By processing first-person video and audio from HoloAssist recordings, the system creates clean audio streams and frames each moment as a decision to either stay silent or provide spoken guidance. Experiments show that EgoVoice improves intervention timing, content relevance, and human preference compared to existing zero‑shot models.
By Heeseung Kim
SpaceCast-Bench is a new benchmark that evaluates predictive spatial reasoning in vision‑language models, moving beyond simple spatial perception to tasks that require constructing scenes, anticipating interventions, and reasoning about unseen outcomes. It contains 3,862 questions from 182 real‑world scenes across 16 task types and three difficulty levels—static perception, local prediction, and global prediction—testing scene understanding, spatial state updating, and relational inference. Evaluation of 21 models shows a large performance gap, with the best model achieving only 58.0% versus 87.2% human accuracy, and fine‑tuning on the benchmark’s programmatically generated data can substantially improve model performance.
By Hongxing Li, Jinyue Su, Dingming Li, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
The paper evaluates the safety of LoopLMs—parameter‑efficient language models that reuse shared parameters across recurrent steps—under jailbreak attacks at various depths. It finds that deeper inference can increase attack success, that safety responses vary across depths, and that attacks transfer between depths, even after standard alignment techniques. To address these gaps, the authors propose SafeBridge, a method that applies depth‑specific control, selective state bridging, and joint safety supervision, which markedly reduces attack success and improves overall utility without increasing over‑refusal.
By Yi Wang, Xiuyuan Qi, Dongqi Han, Dongsheng Li, Wenjie Wang
Large language models often justify legal decisions by citing statutes or precedents, but this study shows that while models correctly name the governing authority 66.7%-100% of the time, the actual verdict rarely changes when the cited authority is swapped—only 0.0%-21.7% on CaseHOLD, 30.0%-76.7% on ECHR and SCOTUS, and 43.3%-50.0% on ContractNLI. This gap persists across models of different sizes and even a purpose‑built legal‑reasoning model. Additionally, a red‑team evaluation revealed that models are more compliant with hidden adversarial instructions than they are sensitive to verdict changes, indicating that citing legal authority is a weak proxy for genuine reasoning and that verdicts remain vulnerable to manipulation.
The paper identifies that visual instance retrieval failures are mainly caused by object-to-image (O2I) ratio mismatch rather than resolution loss, as shown on a benchmark of 3,021 Objaverse objects. Multi-scale architectures reduce resolution effects but still suffer from O2I mismatch, which is asymmetric across query and gallery sizes. The authors propose query-side scale augmentation and an OWLv2 crop reranker, achieving state‑of‑the‑art results on ILIAS 100M, and demonstrate that a LoRA fine‑tune can match these gains in a single forward pass.
The paper introduces Moment-GPT, a tuning‑free pipeline for zero‑shot video moment retrieval that leverages frozen multimodal large language models. It corrects and rephrases queries with LLaMA‑3 to reduce language bias, generates candidate spans using a span generator and MiniGPT‑v2, and selects the best spans with VideoChatGPT and a span scorer. The method outperforms existing MLLM‑based and zero‑shot models on datasets such as QVHighlights, ActivityNet‑Captions, and Charades‑STA.
By Yifang Xu, Yunzhuo Sun, Benxiang Zhai, Ming Li, Wenxin Liang, Yang Li, Sidan Du
The study compares supervised fine-tuning (SFT) and reinforcement learning (RL) as teacher training methods for on-policy distillation (OPD) in a single-teacher setting. Using Qwen3.5-9B models across Agentic, Reasoning, and Perception domains, RL-trained teachers consistently produce stronger students and better recovery of performance gains than SFT-trained teachers. The authors attribute this advantage to RL teachers’ smaller parameter shifts from the shared initialization, which facilitates learning through OPD.
By Jingyuan Huang, Zuming Huang, Yucheng Shi, Zhongzhi Li, Xiaoming Zhai, Wei Chu, Ninghao Liu