The paper investigates on‑policy distillation (OPD) as a preparatory step for reinforcement learning (RL). It shows that students initialized with OPD achieve higher final RL performance than those trained directly with RL or with supervised fine‑tuning followed by RL, even when OPD offers little immediate accuracy gain. The study also finds that the choice of distillation objective (reverse‑KL vs forward‑KL) and the source of trajectories influence OPD’s effectiveness at different stages of RL training.
By Shuai Dong, Yongfu Zhu, Yuqi Xu, Weichu Xie, Liuwenpu, Ziyue Wang, Kaiwen Tuo, Congcong Wang, Siyuan Wang, Wenqi Shao, Shuai Yang, Ji Zhao, Caoyuan Ma, Wenzheng Chang, Taiqiang Wu, Xinlei Yu, Hongrui Wu, Xiaoxuan He, Fangke Chen, Dianyi Wang, Kanghui Tian, Sirry Chen, Xingyu Liu, Xiangnan Wu, Jiawei Guo, Haowen Hou, LingHan Chen, Zhongyu Wei, Jiaqi Wang
The paper investigates how to improve large language model (LLM) based autonomous web agents by filtering irrelevant webpage content. The authors reproduce baseline models on the WebArena benchmark and identify failure modes caused by raw HTML input. They propose DeBERTa‑ and T5‑based retrieval models that rank HTML elements by relevance, fine‑tuned on Mind2Web data, and demonstrate that the DeBERTa model raises the LLaMA‑2‑70B agent’s success rate from 1.97% to 2.96%. Additionally, a zero‑shot ColBERT retriever achieves recall of 0.52 on Mind2Web and 0.47 on WebArena.
By Zhitong Guo, Jing Yu Koh, Ruiyu Li
The study evaluates whether a single large language model (LLM) can handle multiple customer‑support tasks or if separate specialist models are preferable. Using 13 models from five families and 200+ checkpoints across eight datasets, the authors find that multi‑task full fine‑tuning consistently outperforms other strategies. They also show that sequential LoRA and model merging can preserve earlier skills and improve off‑task robustness, offering practical guidelines for real‑world deployment.
By Md Tahmid Rahman Laskar, Xue-Yong Fu, Shashi Bhushan TN
The paper investigates whether using non‑Gaussian priors improves fine‑tuning of large behavior models (LBMs) for robot imitation learning. Across more than 100,000 simulation rollouts and 1,250 hardware trials on diverse tasks, the authors find that non‑Gaussian priors do not yield better fine‑tuning performance than standard Gaussian priors, except possibly at very low data fractions. Diagnostic analyses reveal that encoder training dominates fine‑tuning outcomes, while prior choice has minimal impact.
By Chen Xu, Rishi Shah, Hadas Kress-Gazit, Haruki Nishimura, Masha Itkina
arXiv:2609. 28366v1 Announce Type: cross Abstract: Vision-language models (VLMs) offer a promising approach to long-tail autonomous driving, but existing driving datasets provide limited supervision for connecting decision-critical visual evidence with reasoning and planning.
By Zhipeng Bao, Wenjie Zhao, Tianle Zhu, Haohua Que, Chence Yang, Geng Yuan, Qianwen Li
The paper investigates how fine‑tuning pretrained visual encoders for faithful image reconstruction affects diffusion models that operate in the resulting latent space. It finds that such fine‑tuning reduces the effective dimensionality of the latent representation, causing standard velocity‑prediction flow‑matching to fit noise outside the low‑dimensional signal manifold and making optimization inefficient. Consequently, the authors propose using a clean‑data ($oldsymbol{x}_{0}$) parameterization, which focuses learning on the signal manifold and consistently improves text‑to‑image generation across multiple strong‑reconstruction encoders.
By Chao Feng, Zhiyang Xu, Bowei Chen, Yuanjun Xiong, Xiyao Wang, Jui-Hsien Wang, Richard Zhang, Zhe Lin, Andrew Owens, Yijun Li
Omni-Decision is an omni-modal agent that uses evidence-ledger planning to manage noisy multimodal observations. It replaces the growing dialogue history with a compact evidence ledger that tracks missing, confirmed, and conflicting evidence, allowing the planner to operate on a streamlined context. The system achieves state‑of‑the‑art accuracy on OmniGAIA and WorldSense while operating at a fraction of the cost of larger models.
By Ming Ma, Yi Zhu, Yiran Zhong, Feida Zhu, Yuhao Wang, Junhan Shi, Lingrui Mei, Tianming Yang, Steven Hoi
CAST is a framework for generating anomalous time series that addresses the scarcity and heterogeneity of anomaly data. It uses a two‑stage approach: pretraining on abundant normal data to learn system dynamics, then finetuning with anomaly structure representations to capture diverse anomaly morphologies. Experiments on real‑world datasets show that CAST outperforms existing methods in both generation quality and downstream task performance.
By Haochen Zhang, Jie Peng, Songyuan Sui, Yu-Chao Huang, Xiangqi Zhu, Tianlong Chen
The paper introduces FRESH, a Failure-aware Retrieval framework that uses Experience-Structured Heterogeneous graphs to transform past successes and failures into structured external memory for tool‑using agents. By explicitly modeling dependencies among tasks, actions, errors, repairs, and execution conditions, FRESH enables frozen language models to reuse reliable strategies, avoid recurring failures, and make safer decisions in stateful tool interactions. Experiments on τ‑Bench and AppWorld with multiple open‑source models demonstrate that FRESH consistently improves task success and tool‑use reliability compared to no‑memory agents and other memory‑based baselines.
By Jiaxing Li, Lei Song, Rui Dong, Youyong Kong
Rolling Conformal Prediction (rolling‑CP) is a distribution‑free predictive inference method designed for sequential model training. It calibrates each incoming observation against the current predictor and incorporates it into future training, eliminating the need for data splitting. For exchangeable data, rolling‑CP guarantees marginal coverage with a universal factor‑two bound, and for i.i.d. streams it provides high‑probability training‑conditional validity over time, improving to the target level under stability conditions.
By Chen Cheng, Ruiting Liang, Rina Foygel Barber
The paper examines LLM-based recommendation rerankers that are often evaluated under an oracle protocol, which guarantees the ground-truth item is present in the scored set. Across Amazon datasets, this protocol overestimates realistic NDCG@10 by 92–95% because realistic retrieval only covers 2–19% of relevant items at K=100, creating a recall ceiling that limits any closed-candidate reranker's top‑k NDCG. The authors find that various optimisation strategies—including prompt engineering, model scaling, sequential models, supervised neural rerankers, LoRA fine‑tuning, hybrid retrieval, score‑aware prompting, and LLM+CF fusion—do not significantly improve over a collaborative‑filtering baseline under realistic retrieval, and they propose a Recall‑Aware Evaluation Protocol (RAEP) to better assess rerankers in low‑recall regimes.
By Zhaohui Wang
EvEMTBench is an open, executable, and versioned benchmark designed to standardize the evaluation of machine‑learning methods for power system protection. It defines 12 protection and event‑analysis functions across four grids (20–345 kV) as 24 scored tasks, enabling structured assessment under varied observability, distribution shifts, and cross‑grid transfer scenarios. The benchmark includes committed data partitions, leakage controls, and reproducible reporting, and demonstrates that wider observability does not always help, that shifted conditions expose hidden failures, and that fault detection transfers better than fault localization.
By Julian Oelhaf, Georg Kordowich, Christian Bergler, Andreas Maier, Johann J\"ager, Siming Bayer
The paper introduces a framework that uses world models to enable robotic insertion across diverse parts. By combining proprioceptive data with wrist‑mounted camera visuals, a single model is trained on up to 90 tasks, achieving 56% zero‑shot success on unseen objects versus 7% for a model‑free baseline. The approach scales with more training objects and can be fine‑tuned for improved data efficiency and performance.
By Nicklas Hansen, Iretiayo Akinola, Yijie Guo, Jie Xu, Bingjie Tang, Hao Su, Xiaolong Wang, Abhishek Gupta, Dieter Fox, Yashraj Narang
The paper introduces Causal Shortcut Learning (CSL), a framework that identifies token chains—called causal shortcuts—that guide Diffusion Language Models (DLMs) toward correct reasoning paths. By extracting these shortcuts and applying parallel prioritized masking during training, CSL improves both convergence speed and generation accuracy. Experiments on several reasoning benchmarks and two base models show CSL outperforms existing SFT-variant baselines, achieving an average 1.92% improvement over SFT-only models and up to 4.20% on MATH-500.
By Dian Jin, Kairong Han, Baohong Li, Xinpeng Dong, Zijing Hu, Nuanqiao Shan, Fei Wu, Kun Kuang
The paper argues that meaning identity—whether two sentences convey the same idea after wording changes—is not encoded in the geometry of independently produced sentence embeddings. Experiments on frozen off‑the‑shelf encoders and language models show that identity can only be reliably computed when both sentences are processed together in a single forward pass, yielding high accuracy (0.90–0.96) on PAWS‑X, whereas independent embeddings or simple fusion methods perform near chance. Even advanced bi‑encoder fine‑tuning improves performance on PAWS but fails to generalize to other similarity tasks, underscoring that identity is a cheap computed operator rather than a property of individual sentence vectors.
By Jiaqi Deng
RapidUn is a parameter reweighting framework that uses influence estimates to guide LoRA-only updates for efficient unlearning of targeted behaviors in large language models. It operates in a practical PEFT setting with a small forget set and limited retain buffer, converting cross-sample influence into fixed sample-specific weights for weighted LoRA unlearning. Experiments on Llama‑3‑8B with Dolly‑15k and Alpaca‑57k datasets show RapidUn achieves lower trigger ASR than Fisher, GA, and LoReUn while preserving clean utility, and delivers a 77× wall‑clock speedup over clean‑corpus LoRA retraining, with additional evaluations supporting its effectiveness.
By Guoshenghui Zhao, Huawei Lin, Weijie Zhao
The paper presents a data‑centric approach to improve automatic speech recognition for non‑verbal vocalizations (NVVs) in the ISCSLP NVVSpeech Challenge. It introduces cross‑dataset label harmonization and a two‑stage sampling schedule—first square‑root category sampling to address long‑tailed distributions, then uniform‑category fine‑tuning—to jointly transcribe lexical content and 16 NVV categories. The final system achieved an official score of 63.86, ranking fourth in Track 1.
By Shangyue Jia, Jingru Ma, Yangzhuo Li, Daoping Luo, Bowen Tian, Hanchen Lu, Wenze Ren, Yunxiang Chen, Houdun Liu, Su Feng, Lei Xie, Liumeng Xue
The paper introduces "open‑set adversarial forgetting" to make selected identities unlinkable in face recognition models while keeping the system functional for other users. Three loss functions are proposed: one disperses embeddings from their centroid, and two map each image to a near‑orthogonal target—either learned via the classifier head or fixed as an almost‑orthonormal frame. Experiments on multiple backbones and forget scales show that these losses, especially the orthonormal‑frame approach, effectively render the targeted identities unidentifiable and outperform prior methods while preserving higher retain rates.
By \"Unsal \"Ozt\"urk, Vedrana Krivoku\'ca Hahn, Sushil Bhattacharjee, S\'ebastien Marcel
arXiv:2609.27987v1 Announce Type: new
Abstract: Large language models (LLMs) have made substantial progress on medical question-answering, yet effective medical dialogue also requires learning to ask...
By Chenxuan Li, Jiayi Wan, Xinrong Chen, Zhongyu Zhao, Xuecheng Shang, Peixing Wan
arXiv:2609.27811v1 Announce Type: cross
Abstract: Online platforms have become arenas for the public contestation of climate change, shaping how scientific knowledge, denial, and uncertainty are expr...
By Daniel Morais, Diego H. M. Magalhaes, Gabriel H. Silva, Andrea Failla, Valeria de C. Santos, Helen C. S. C. Lima, Carlos H. G. Ferreira