The paper investigates model merging as a scalable alternative to multi-task training for multi-domain automatic speech recognition (ASR). It benchmarks 11 merging algorithms across 10 European Portuguese domains, assessing in‑domain accuracy, robustness to distribution shift, and performance on English and multilingual data. The authors introduce BoostedTSV‑M, a new merging method that mitigates rank collapse through singular‑value boosting, achieving better performance than full fine‑tuning while maintaining out‑of‑distribution generalisation in a single model.
By Carlos Carvalho, Francisco Teixeira, Thomas Rolland, Alberto Abad
RECAST is a reconstruction-based framework that generates chain-of-thought (CoT) explanations from closed‑source reasoning models that only expose final answers, such as Claude‑Sonnet‑5.5, GPT‑6.1, and Gemini 3. Unlike prior extraction methods that rely on visible reasoning fields or scratchpads, RECAST conditions a decoder on the question and the model’s final answer to produce a first‑person structured rationale, cleans it, and outputs data ready for supervised fine‑tuning. The system is open‑source, packaged as an agent skill, and demonstrates successful qualitative reconstructions on the mentioned models using an answer‑only protocol.
By Qi Wang, Kang Xu, Chengcheng Wan, Jiangtao Wang
The study evaluates whether a large language model (LLM) can manage multi‑zone variable‑air‑volume HVAC control without building‑specific training, and whether reinforcement fine‑tuning verified by rollout can enhance a smaller, deployable model. Five controllers were tested in a physics‑based four‑zone emulator over five summer days; the LLM (GPT‑5) reduced electricity use by 6.8‑7.4% compared to a guideline baseline but performed worse on outdoor‑air usage and CO₂ compliance. A reinforcement‑learning controller achieved a 4.5% electricity savings while improving temperature and CO₂ compliance, yet its critic’s high correlation with deterministic rollout did not reliably rank actions, and fine‑tuning a smaller LLM with rollout returns yielded no sustained performance gains due to unreliable next‑state predictions.
By Takumi Shioda, Kohei Terashima, Tatsuo Nagai
The paper introduces Counterfactual Pairwise Augmentation (CPA), a method that enhances small language model judges by generating counterfactual responses to clarify decision boundaries during training. CPA jointly fine‑tunes models on both pointwise and pairwise data, enabling the judge to operate in pointwise mode while benefiting from explicit pairwise comparisons. Experiments on Qwen3‑8B adapters show CPA improves macro‑F1 scores by 2.5 points over pointwise‑only training and outperforms both model merging and the Gemini 3.6 Flash zero‑shot judge by significant margins.
By Mohammad Zangooei, Peyvand Teymoori, Mohammad Robati Shirzad, Soroosh Shahtalebi
The paper presents a multi‑task neural network and scalar‑potential solver for reconstructing bilateral coronary artery trees from cardiac CT angiography. It operates on $128^3$ regions of interest, using a 3D pyramid encoder, tubular decoder, and specialized heads for sub‑voxel displacement, ipsilateral territory prediction, and ostium localization. The method achieves sub‑millimetre accuracy on out‑of‑sample and development datasets, outperforming baseline 3D U‑Net models.
By Zeyu Jia, Xin Ming
Latent‑Lens is a lightweight codec that enables a frozen vision‑language model (VLM) to communicate visual information to a frozen small language model (SLM) directly in latent space, avoiding the lossy bottleneck of discrete text. The codec uses a domain encoder to project VLM hidden states into the SLM’s embedding space and an attention decoder (a LoRA) that teaches the SLM to attend to these visual prefixes, training only about 1.6 M parameters (0.2 % of the combined model size). On visual question answering tasks (CLEVR, GQA) and image captioning (Flickr8k), Latent‑Lens significantly outperforms text‑based communication with the same LoRA budget, achieving up to +69.8 pp improvement on VQA and +18.8 pp on captioning.
By Mo Shahdloo, Faegheh Sardari, Mathew Salvaris
The paper formalizes controllable user simulation as a causal inference problem, showing that standard supervised fine‑tuning on post‑hoc trajectory labels introduces a look‑ahead bias that breaks causal consistency. This bias leads to a phenomenon called controllability collapse, where evaluation metric variance explodes under policy shift. The authors propose theoretical conditions for accurate simulation and practical training mitigations—such as a priori controls, step‑wise dynamic controls, and direct policy‑conditioned learning—to eliminate look‑ahead bias, preserve natural variance, and achieve robust zero‑shot generalization to unseen agent behaviors.
By Guy Tennenholtz, Ofer Meshi, Amir Globerson, Uri Shalit, Jihwan Jeong, Craig Boutilier
The study investigates how Jacobian-based fine‑tuning affects neural emulators of chaotic partial differential equations (PDEs). By decomposing error growth into physical and model‑induced components, the authors introduce an amplification ratio that quantifies how the emulator’s dynamics magnify prediction errors. Across 21 experimental setups, Jacobian fine‑tuning significantly reduced errors—especially for configurations with high amplification ratios—while matching the solver’s Jacobian along the model’s error direction (EDJM) outperformed random directions and input‑noise injection.
By Shreeyam Bangera, Manya Gupta
The paper investigates a new security vulnerability in fine‑tuned Mixture‑of‑Experts large language models, termed structural memorization, where the model not only memorizes training data but also the internal routing pathways used during fine‑tuning. The authors demonstrate that these memorized expert pathways can be recovered from the model’s unconditional generation using only routing and output logits, even without access to the fine‑tuning data. They further show that this vulnerability is amplified in distributed deployments and that traditional membership inference metrics fail to detect it, proposing composite metrics that improve detection of memorized routing structures.
By Fei Wang, Baochun Li, Bo Li
TalentRL is a strategy‑guided reinforcement learning framework that routes natural‑language skills hierarchically, using coarse‑grained talents to guide fine‑grained skill retrieval. The method trains talent routing via reward contrasts and optimizes execution with group‑relative advantages, achieving significant performance gains on open‑domain QA benchmarks and agentic tasks such as ALFWorld and WebShop.
By Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang, Michelle Chang, Jingwen Lu
The paper introduces LOCOGEN, a method for designing regulatory DNA sequences that balances activity, novelty, and plausibility. LOCOGEN treats guidance as a locality problem, protecting motif positions with an invariant mask while applying out‑of‑distribution damping only to background regions. Experiments on HepG2, K562, and SK‑N‑SH datasets show that LOCOGEN reaches a previously underexplored region of the property‑novelty‑plausibility space, outperforming existing baselines.
By Dabin Jeong, Marwane Bourdim, Mo Lotfollahi
Mage‑VL‑Cowork is a 4‑billion‑parameter hybrid computer‑use agent that integrates visual GUI interaction with code and terminal execution in a single policy. It is trained via a progressive five‑stage curriculum—mid‑training, supervised fine‑tuning, and reinforcement learning—starting from a multimodally aligned checkpoint. The training pipeline, Mage‑CUA‑Forge, automates task generation and skill conditioning, producing nearly 300,000 desktop and web tasks while cutting per‑task costs by 82.6%, and the resulting model achieves 66.0% on OSWorld‑Verified, setting a new state‑of‑the‑art for its size.
By Xiaoyi Zhang, Haoqing Wang, Xinjie Zhang, Xiang Ming, Xing Wang, Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Andong Tan, Xinjiang Wang, Jinghao Guo, Yifei Shen, Xiao Li, Peng Zhang, Xun Guo, Jiahao Li, Shicheng Zheng, Jinglu Wang, Zongyu Guo, Wenxuan Xie, Bin Li, Yan Lu
Steering vectors (SVs) are residual-stream directions that can shift LLM behavior when added to hidden states during inference. This study investigates whether the same SVs can guide parameter updates during post‑training, exploring supervised learning (SFT and distillation) and reinforcement learning with verifiable rewards (RLVR). The authors find that SV injection improves performance over plain SFT and distillation, achieving up to a 12% gain in certain distillation settings, and that applying a correctness‑contrast direction as advantage repair in RL recovers learning signals for held‑out math benchmarks. They also identify failure boundaries related to code generation, cross‑family layer transfer, and inference‑time application on SV‑trained checkpoints.
By Tiejin Chen, Maunil Vyas, Huaiyuan Yao, Hua Wei
The paper introduces PG‑FM, a Prompt‑Guided Feature Modulation method that adapts RGB‑thermal object detectors to the estimated corruption type and severity of each modality. Unlike restoration‑based approaches, PG‑FM performs per‑input adaptation without significant latency and maintains clean‑set performance while boosting mAP50 by up to 20.8 points on FLIR and 19.9 points on KAIST under simulated corruptions. The method also transfers effectively to unmodified, naturally acquired external RGB‑T datasets without retraining.
By Liuyin Shi, Chengqi Li, Wenbo He
The paper investigates where fine‑tuned behaviors in language models are instantiated by patching the key‑value (KV) cache between a LoRA‑fine‑tuned model and its base version. Results show that certain behaviors, such as German responses or emergent misalignment, can be transferred or removed by swapping KV cache segments, and that the impact grows with the amount patched and is concentrated at a few pivotal positions like system‑prompt and chat‑template tokens.
By Jonas Loos, Oliver Eberle
arXiv:2610. 41588v1 Announce Type: cross Abstract: Direct Preference Optimization (DPO) treats every reported comparison as correct, yet preference data contain reversed, tied or otherwise noisy comparisons.
By Bin Luo, Param Venkat Vivek Kesireddy, Anil Vallepu
The paper introduces BSTRB, a benchmark of 8,398 noisy image samples with high‑fidelity annotations from 1,699 real‑world video sequences, designed to evaluate Vision‑Language Models (VLMs) on degraded text. It presents Hawkeye, a VLM fine‑tuned on BSTRB that reads and reasons about illegible text by using contextual priors and pixel‑level alignment. Hawkeye performs comparably to larger VLMs overall and surpasses them on text at or below 20 pixels, making it suitable for real‑time deployment in noisy egocentric settings.
By Brian Tang, Kang G. Shin
The paper studies parameter‑efficient fine‑tuning (PEFT) methods—LoRA and QLoRA—on pretrained music encoders for audio, sheet‑music images, and symbolic scores. It compares full fine‑tuning with PEFT across nine encoders on TriScore and the Musical Theme Dataset, finding that LoRA/QLoRA often match or surpass full fine‑tuning while using far fewer parameters. However, gains in unimodal discrimination do not consistently translate to improved multimodal retrieval, though the adapted-and-projected representations achieve strong piece‑level retrieval performance on TriScore.
By Eliseo Fuentes-Martinez, Antonio Hidalgo-Centeno, Jorge Calvo-Zaragoza, Antonio Javier Gallego
The paper introduces Latent Reasoning in Continuous space (LARC), a framework that replaces discrete token-based reasoning with reasoning over continuous hidden states to better align with image generation’s continuous nature. LARC is trained in two stages: curriculum supervised fine‑tuning that gradually shifts from token to latent reasoning, followed by an information‑gain reinforcement learning stage that refines latent reasoning without ground‑truth supervision. Experiments on multiple text‑to‑image benchmarks show that LARC improves generation quality, prompt adherence, and compositional fidelity, achieving state‑of‑the‑art results on GenEval and demonstrating robust test‑time scaling.
By Byungwoo Jeon, Yoonwoo Jeong, Hyunseok Lee, Minsu Cho, Jinwoo Shin
The paper investigates how data augmentation can improve sentiment classification on small datasets. Three augmentation techniques—Easy Data Augmentation (EDA), back‑translation, and contextual token substitution—are evaluated with both traditional classifiers (Logistic Regression, Random Forest) and BERT. Results show that all methods boost performance, with contextual augmentation best for BERT and EDA/back‑translation benefiting traditional models.
By Mahmoud S. Mayaleh, Samer A. Mayaleh