The paper introduces SBERT2S1, a method that transforms Sentence-Transformers encoders into typed decision models for biomedical text, and presents BIODECIDE, a suite for evaluating such models, along with MEDLINE‑S1, a large training set of 243k decisions derived from NLM indexing. Experiments show that retrieval‑trained encoders improve zero‑shot matching of content‑bearing options, and that a prior‑fused residual (PFR) head benefits most from retrieval pre‑training, while a cross‑head (C) head generally outperforms PFR across all objectives. The authors also release code, data, and a model, and discuss calibration and reward‑normalisation effects in their RLCD recipe.
By Pritam Deka
The paper presents a method for improving cross‑lingual retrieval in multilingual encoders by adding a query‑only low‑rank adapter and deterministic routing based on language. The adapter, trained against frozen document embeddings, is used for cross‑lingual queries while the original encoder handles same‑language queries, preserving existing performance. Using the SampoTron LoRA adapter with the Nemotron‑3‑Embed‑1B model, the authors report a 20.9% relative gain in nDCG@10 across six English, Finnish, and Swedish directions on a financial benchmark, with all cross‑lingual directions improving and same‑language performance maintained.
By Akshay Jain, Edward Kim
The paper investigates how enrichment—adding intermediate guidance to sparse rewards in reinforcement learning—affects policy-gradient updates. It shows that omitting or truncating importance-weight corrections implicitly reweights rewards, and decomposes the resulting gradient error into scale, rotation, and variance components. The authors propose a sequential Monte Carlo weight correction that mitigates variance growth, and apply their analysis to a fine-tuned Qwen3-1.7B model on OpenMathReasoning, illustrating how enrichment helps prevent collapse in sparse domains.
By Jinwoo Kim, Shraddha Barke
The paper investigates how training large language models to use fewer tokens in chain-of-thought (CoT) reasoning impacts the faithfulness and monitorability of the generated explanations. Three efficiency methods—fixed generation budget, per-example length target, and group-relative length reward—were applied during fine‑tuning, and the resulting models were evaluated on how well their CoT reflects decision processes and whether it signals changes due to input interventions. Results show that while faithfulness generally decreases because models become less consistent, monitorability remains relatively robust, with models still indicating the influence of input changes even when CoT is shortened.
By Samuel Lewis-Lim, Xingwei Tan, Mario Sanger, Zhixue Zhao, Nikolaos Aletras
The paper introduces Inner Momentum (IM) for differentially private Muon training. It addresses distortion caused by per-example gradient clipping by averaging each example’s Muon gradient over the current and recent models before clipping, thereby bounding clipping-induced distortion. Experiments on private GPT‑2 fine‑tuning show that DP‑Muon‑IM consistently improves BLEU and ROUGE‑L scores and reduces polar error compared to standard DP‑Muon.
By Bishnu Bhusal, Minh Vu, Ben Southworth, Geigh Zollicoffer, Rohit Chadha, Manish Bhattarai
The paper introduces Variational Risk Minimization (VRM), a distillation framework that treats large vision‑language model (LVLM) report variants as Monte Carlo samples of latent clinical interpretations. VRM learns from a variationally marginalized teacher distribution, providing uncertainty‑aware supervision even when modalities are missing, and outperforms fine‑tuning baselines while improving calibration. In a compact edge‑student setup, a confidence‑gated cascade achieves an AUC of 0.941 at 103 ms latency with only 20.3 % cloud escalation, offering a clear reliability‑latency trade‑off for cloud‑edge medical imaging workflows.
By Xinye Yang, Zhusi Zhong, Scott Collins, Michael Bernstein, Grayson Baird, Terrence Healey, Michael Atalay, Mahesh Jayaraman, Xuyu Wang, Zhicheng Jiao
The paper investigates how to adapt large language models to mimic an individual author's style using only a few example abstracts, a task made harder by the formal nature of scientific writing. It proposes three style‑conditioning methods—contrastive activation steering, a network predicting steering vectors, and a hypernetwork predicting LoRA adapters—and finds that while fine‑tuning captures the strongest style signal, it harms fluency; the hypernetwork offers the best balance between style imitation and output quality for both seen and unseen authors. The authors also show that steering can be performed at the author level by contrasting author abstracts against style‑neutral generations for the same content, eliminating the need for a predefined style inventory and outperforming inventory‑based approaches.
whyItMatters":"The study provides practical techniques for author‑style transfer in scientific writing, revealing a trade‑off between style fidelity and fluency and demonstrating that hypernetworks can effectively balance these aspects."
By Leonard Popp, Danni Liu, Supriti Sinhamahapatra, Jan Niehues
The paper proposes a method for image-based traversability estimation that uses von Mises‑Fisher mixture prototypes in a frozen vision‑language feature space. By incorporating relative natural‑language rules as a commonsense prior and fine‑tuning with sparse image annotations, the approach achieves domain adaptation with fewer labels. Experiments on WayFAST show competitive accuracy to end‑to‑end models and improved dense predictions on semantic maps, while qualitative results illustrate zero‑shot language prior use and semantic interpretation of prototypes.
By Simon Schwaiger, David Seyser, Alessandro Scherl, Zlatan Ajanovi\'c, Wilfried W\"ober, Gerald Steinbauer-Wagner
The paper presents a system for the Touché 2026 causality extraction challenge, focusing on counter‑causal claims—sentences that appear causal but actually deny causation. It tackles three subtasks: detecting causal sentences, extracting cause and effect spans, and labeling polarity (procausal, counter‑causal, or uncausal). The approach uses a fine‑tuned classifier with a cross‑task rule for detection, an ensemble of three RoBERTa‑large BILOU+CRF taggers for extraction, and counter‑causal data augmentation via a large language model for polarity classification, achieving state‑of‑the‑art scores on the CCNC test set.
By Roham Zendehdel Nobari, Shayan Sooratgar
The paper presents a system for matching job candidates to vacancies that provides interpretable evidence rather than a single relevance score. It uses a two‑stage approach: an LLM‑based labeler refined through recruiter feedback and a distilled bi‑encoder that runs online on CPU. The model, trained on 168,772 labeled pairs, achieves 95.79% agreement with recruiter‑recorded decisions on a production‑feedback subset.
By Ilya Chekin (BroutonLab), Vyacheslav Malyugin (BroutonLab), Vladimir Chirkov (BroutonLab), Mikhail Yurushkin (Curately)
The paper introduces a self‑supervised framework called software‑in‑the‑loop reconstruction (SWR) that automatically generates reference outputs and verification targets for terminal agents by leveraging existing scientific software workflows. SWR executes multiple input configurations, partitions cases into public observations and hidden evaluations, and uses a hierarchical verifier to assess agent‑generated programs against hidden workflow outputs. The authors demonstrate the approach on 500 workflows across six domains, achieving significant performance gains on the Terminal‑Bench 2 benchmark with the Qwen3.8‑Max and Qwen3.8‑27B models.
By Zhongzhi Li, Yucheng Shi, Zongxia Li, Junyao Yang, Ruhan Wang, Yu Wang, Jingyuan Huang, Jichao Yu, Ninghao Liu, Haitao Mi, Leowei Liang
The paper proposes using Small Language Models (SLMs) as a task‑tool relevance classifier to verify each tool call made by AI agents. By evaluating every selected tool against the assigned task, the SLM provides a relevance signal that can be used for downstream enforcement. The authors introduce a novel dataset of multi‑tool tasks across distinct Model Context Protocol servers and explore prompt‑optimization, supervised fine‑tuning, and reinforcement learning (GRPO) to optimize and specialize the SLMs.
By Chiara Troiani, Arash Salarian, Majed El Helou, Benjamin Ryder, Jean Diaconu, Herv\'e Muyal, Marcelo Yannuzzi
The paper introduces a hybrid BERT‑CNN‑BiLSTM model for simultaneously classifying Bangla news headlines and performing sentiment analysis. Using the BAN‑ABSA dataset of 9,014 headlines, the authors compare two sampling strategies and report headline accuracy up to 81.37 % and sentiment accuracy up to 73.43 %. The proposed framework outperforms baseline models and establishes new state‑of‑the‑art results for Bangla headline classification and sentiment analysis.
By Mirza Raquib, Munazer Montasir Akash, Tawhid Ahmed, Saydul Akbar Murad, Farida Siddiqi Prity, Mohammad Amzad Hossain, Asif Pervez Polok, Nick Rahimi
The paper introduces SafeCut, a method for Source-Free Domain Adaptation that uses Vision‑Language models as external knowledge. SafeCut employs a cut statistic to gauge prediction reliability, enabling a dynamic, reliability‑gated supervision between the source‑pretrained model and the ViL model. This approach selectively amplifies correct mutual corrections while suppressing error propagation, achieving state‑of‑the‑art performance on multiple SFDA benchmarks.
By Seongjun Lee, Changhee Lee
The paper investigates why supervised fine-tuning (SFT) causes more forgetting than reinforcement fine-tuning (RFT) on classification tasks where tokens share a semantic class but differ in style. By deriving a linear-softmax policy decomposition, the authors show that SFT and RFT have parallel semantic updates but diverge in style dynamics, with SFT exhibiting off‑axis style drift that leads to a positive lower bound on forgetting. Theoretical results are supported by simulations over task sequences.
By Haodong Liang, Yanhao Jin, Krishnakumar Balasubramanian, Lifeng Lai
The paper evaluates two strategies for training language‑model‑based world models—fine‑tuning and Retrieval Augmented Generation (RAG)—across five diverse environments. Fine‑tuning generally yields higher rewards, while RAG is more data‑efficient. The authors propose a counterfactual error estimation for RAG, improve retrieval with a hierarchical query reformulation, and combine both approaches into a hybrid system that outperforms each method alone.
By Dhananjay Ashok, Shantanu Agarwal, Vivek Datla, Jonathan May, Alfy Samuel
The paper introduces a de‑identified corpus of 33 in‑person cognitive assessment conversations, comprising 8,250 utterances annotated for three speaker roles and 56 dialogue acts. The authors benchmark large language models on fine‑grained dialogue‑act classification and next‑patient‑utterance generation, finding that instruction tuning and reasoning‑aware fine‑tuning improve performance but that models still struggle with closely related dialogue acts. The corpus and benchmark are presented as tools to measure interaction structure in cognitive assessments and to support future research on conversational markers, clinician education, and validated simulated patients.
By Vishalakshi Arumugam, Dan Schumacher, Veronica Rammouz, Erfan Nourbakhsh, Enrique Gonzalez Guerrero, Jeremy Davis, Anthony Rios
The paper introduces Follow the Winners (FTW), a critic‑free policy‑learning algorithm that adapts the cross‑entropy method for reinforcement fine‑tuning of agentic large language models. FTW replaces group rollouts with an ordinal filter on replay‑buffer samples, achieving polynomial concentration in the order statistic of returns and offering a bounded risk‑seeking offset that balances variance reduction. Experiments on Sokoban and Search‑R1 show that FTW matches the performance of GRPO and PPO while reducing reliance on value models or repeated rollouts.
By Joery Ari\"en de Vries, Neil David Lawrence, Zhenwen Dai
Multimodal representation learning has largely relied on contrastive models like CLIP that produce a single embedding per sample, which limits their ability to capture relation-dependent relevance. The proposed Relation-Conditioned Multimodal Learning (RCML) framework explicitly conditions embeddings on natural‑language relation descriptions, enabling the same sample to be represented differently under various relational contexts. RCML builds relation‑aware training pairs, incorporates a relation‑conditioned module, and uses a unified contrastive objective to jointly model cross‑modal alignment and relation‑induced structure, achieving superior performance on retrieval and classification tasks across zero‑shot, fine‑tuned, and out‑of‑domain settings.
By Yang Qiao, Yuntong Hu, Bowen Zhu, Hasibul Haque, Liang Zhao
The paper introduces checkpoint handoff, an evaluation protocol that separates an agent’s ability to reach useful states from its ability to complete tasks in reinforcement learning. By using one checkpoint as a reacher up to a handoff point and another as a solver from the same replayed history, the authors can measure Reach (how often states within a fixed number of actions from success are achieved) and Solve (how often the task is completed from those states). Experiments on TravelPlanner and ALFWorld show that switching the solver from supervised fine‑tuning to RL yields larger gains when RL is used as the reacher, indicating that RL more effectively finds solvable states.
By Xuan Liu, Jingbin Qian