The paper examines how large language models (LLMs) can be biased by irrelevant social contexts when evaluating teachers, using a large U.S. classroom transcript dataset. It shows that spurious contexts can shift model ratings by up to 1.48 points on a 7‑point scale and that standard mitigation methods like SFT and DPO are insufficient. The authors introduce Debiasing‑DPO, a method that combines contrastive reasoning‑augmented DPO with SFT, which reduces bias by 84% and improves predictive accuracy by 52% on Llama and Qwen Instruct models.
By Hyunji Nam, Dorottya Demszky
The paper introduces Paraesthesia, a dynamic backdoor attack that uses emotionally styled inputs as triggers for large language models. By mapping target emotions into a valence–arousal space and rewriting a small subset of clean samples, the attack achieves over 98% success while minimally affecting clean performance. Experiments on four major LLMs show that the trigger cannot be fully explained by token-level cues and remains robust against several filtering and mitigation techniques.
By Ziyu Liu, Tao Li, Tao Yang, Tianjie Ni, Xiaolong Lan, Wengang Ma, Junjiang He
The paper introduces a scalable product‑linking system that uses a retrieve‑then‑match cascade. First, a lightweight text cross‑encoder auto‑resolves the majority of merchant‑catalog product pairs with high precision, while an agentic multimodal vision‑language model handles the remaining ambiguous cases by inspecting images and performing web searches. This approach balances computational cost and accuracy, improving overall link coverage from 68% to 77% without requiring fine‑tuning of the agent.
By Jian Wang, Steven Xu, Sanjyot Thete, Maryam Barouti, Tom Tang, Elaine Wu, Charu Sareen, Kyle MacDonald
The paper investigates how parameter‑efficient fine‑tuning affects a language model’s ability to learn new facts while retaining existing knowledge. Using a controlled OpenStreetMap benchmark, the authors compare full fine‑tuning with quantized low‑rank adaptation (QLoRA) at various ranks, finding a clear trade‑off: lower ranks preserve out‑of‑domain performance but acquire fewer facts, whereas higher ranks improve factual generalization at the cost of unrelated benchmark performance. Diagnostics in distribution, weight space, and spectral analyses confirm this acquisition‑retention frontier, and a separate math adaptation experiment shows a weaker effect, indicating the phenomenon is strongest when installing new factual associations.
By Estelle Zheng, S\'ebastien Warichet, Emmanuel Helbert, Christophe Cerisara
The paper introduces SalesLLM, a bilingual (Chinese/English) benchmark for evaluating large language models (LLMs) in realistic sales dialogues. It comprises 30,074 scripted configurations and 1,805 curated multi‑turn scenarios from Financial Services and Consumer Goods, with controllable difficulty and personas. An automatic evaluation pipeline uses an LLM judge for sales‑process progress and fine‑tuned BERT classifiers for end‑of‑dialogue buying intent, while a user model, CustomerLM, is trained to improve simulation fidelity. SalesLLM scores correlate strongly with human ratings (Pearson r = 0.86) and reveal that top Chinese LLMs match junior‑to‑intermediate human salespeople but not experts, with cross‑lingual consistency remaining poor.
By Xuanbo Su, Wenhao Hu, Le Zhan, Yuting Xie, Kailin Lyu, Kaijie Chen, Ziwei Li, Yeqiang Wang, Haibo Su, Yunzhang Chen, Ling Huang
The paper introduces CoDIT, a contrastive decoding technique that separates instruction-following behavior from pre-trained world knowledge in large language models. By generating responses that emphasize post-training instruction capabilities while suppressing shared pre-trained knowledge, CoDIT creates instruction-tuning datasets that lead to consistently better model performance than directly generated responses or existing public datasets. The authors also provide theoretical and empirical evidence that CoDIT effectively distills instruction-tuning knowledge from model parameters into text, facilitating cross-architecture transfer.
By Tatsuya Ichinose, Youmi Ma, Masanari Oi, Ryuto Koike, Naoaki Okazaki
The paper investigates how refusal training shapes the internal geometry of language models, showing that activation updates from refusal-completion losses create a distinct low‑dimensional refusal subspace. In a case study on OLMo‑2‑0425‑1B‑Instruct, the authors link the brittleness of refusal directions to repetitive refusal prefixes and demonstrate that using diverse refusal starts can increase the stable rank of gradients, thereby hardening the model against vector‑ablation attacks. The work provides insights into the emergence of safety‑critical features and offers a potential strategy to strengthen refusal robustness.
By Andrey Labunets
The paper introduces FairTPT, a fairness-aware test‑time prompt tuning method for vision‑language models like CLIP. It jointly minimizes target marginal entropy while maximizing spurious marginal entropy to reduce bias under subpopulation shifts. Experiments show that standard episodic test‑time adaptation can worsen disparities, but FairTPT outperforms existing debiasing methods while preserving overall performance.
By Yoann Launay, Parameswaran Kamalaruban, Tom Kempton, Stuart Burrell, David Sutton
MeMark introduces a watermarking scheme for Spiking Neural Networks that embeds a multi‑bit identifier directly into the membrane state of selected Leaky Integrate‑and‑Fire neurons, rather than in the output head. The watermark is recoverable by comparing neuron firing thresholds, eliminating the need for a learned decoder. Experiments on various SNN architectures—including a 215.4M‑parameter SpikeGPT checkpoint—show that all 20 independent 64‑bit keys reliably pass verification under a 51/64 rule, remain robust after fine‑tuning, pruning, quantization, and output‑head replacement, and are not recovered by random keys or adaptive attacks within the tested threat model.
By Roberto Ria\~no, Gorka Abad, Stjepan Picek, Aitor Urbieta
The paper applies sparse autoencoders to a neutrino foundation model trained on IceCube data, uncovering a validated atlas of physical concepts within the model’s internal representation. Causal analysis shows the direction reconstruction head largely ignores this atlas, whereas an uncertainty head trained on the same representation effectively uses quality and brightness features, improving angular resolution from 20.2° to 3.2° at 20% efficiency. These findings demonstrate that mechanistic interpretability can expose latent physics and guide the design of downstream tasks.
By Rapha\"el Bonnet-Guerrini, Johann Ioannou-Nikolaides, Inar Timiryasov, Vincenzo Piuri
The paper introduces COMLLM, a generative framework that combines Group Relative Policy Optimization with a Look‑Ahead Collaborative Simulation to enable multi‑turn reasoning for task offloading in Mobile Edge Computing. By performing multi‑step Monte Carlo rollouts that jointly model server queue dynamics, COMLLM incorporates long‑term system evolution into its reward design, achieving near‑optimal latency and improved load‑balancing fairness. The framework demonstrates zero‑shot scalability to larger network topologies, outperforming supervised fine‑tuning, deep reinforcement learning, and heuristic baselines without requiring retraining.
By Ning Yang, Chuangxin Cheng, Haijun Zhang
The paper introduces a minimalist visual-inertial odometry system that uses only four downward-facing photodiodes with optical Gabor masks and an IMU to estimate motion for differential-drive robots. By jointly optimizing mask parameters and a Temporal Convolutional Network in a physically-grounded simulator, the model decodes speed from the photodiode signals and combines it with IMU angular speed to produce a continuous planar trajectory. Experiments on a prototype robot across indoor and outdoor terrains show that the system closely follows reference trajectories without real-world fine-tuning.
By Francesco Pasti, Jeremy Klotz, Nicola Bellotto, Shree K. Nayar
The study fine‑tunes the Whisper Small model for Automatic Speech Recognition (ASR) in Baniwa, an indigenous Arawakan language. Using a 0.54‑hour corpus of 1,373 manually transcribed recordings, the fine‑tuned model achieved a Word Error Rate of 37.5% and a Character Error Rate of 7.45%. These results provide an initial baseline for Baniwa ASR and suggest that multilingual foundation models can be adapted to extremely low‑resource languages.
By Leonardo Duart, Tiago Fonseca, Thiago Chac\'on
LibriBrain100 is a new large‑scale MEG dataset for speech decoding that contains over 100 hours of high‑quality recordings while subjects listened to naturalistic continuous speech. The dataset more than doubles the size of the original LibriBrain release, with a record 80 hours from a single subject and additional 40‑minute recordings from 32 subjects. The authors demonstrate the value of deep within‑subject data and broad multi‑subject data by achieving state‑of‑the‑art word‑classification performance and showing that supervised fine‑tuning can compensate for limited per‑subject data, all supported by open‑source tools and a public competition leaderboard.
By Francesco Mantegna, Dulhan Jayalath, Gereon Elvers, Tasha Kim, Benjamin Ballyk, Alex Fung, SungJun Cho, Teyun Kwon, Luisa Kurth, Miran \"Ozdogan, Gilad Landau, Pratik Somaiya, Natalie Voets, Mark Woolrich, Oiwi Parker Jones
The paper introduces Code World Model, a framework that decouples world evolution from visual rendering by using a coding agent as a world brain. The agent reasons about events, generates executable code to maintain persistent state, and a proxy representation links this state to a video model for high‑fidelity visual output. Experiments with MiniMax‑H3 show that the system can follow proxy‑based spatiotemporal specifications while preserving rich visual dynamics, illustrating a new approach to open‑ended world modeling.
By Yiwen Chen, Guosheng Lin, Chi Zhang
The paper introduces Efficient Retrieval Adapter (ERA), a query‑side adapter framework that enables dense retrieval systems to adapt to asymmetric query–document scenarios without re‑indexing. ERA first aligns the embedding spaces of a powerful query embedder and a lightweight document embedder using unlabeled documents, then fine‑tunes the aligned query representation with a small set of labeled query‑document pairs. In experiments on 126 MAIR retrieval tasks across six domains, ERA boosts average nDCG@10 by up to 8.2 points in symmetric settings and over 12 points in asymmetric settings while requiring far fewer labels than fully supervised adapter training.
By Seiji Maekawa, Moin Aminnaseri, Pouya Pezeshkpour, Estevam Hruschka
The paper introduces MAMA-FLUX.2, a conditional latent flow‑matching model built on FLUX.2-Klein-4B, designed to synthesize post‑contrast breast DCE‑MRI from pre‑contrast images for the MAMA‑SYNTH challenge. It encodes the pre‑contrast image as spatial conditioning and predicts the flow field for the post‑contrast latent, adapting the pretrained model with LoRA fine‑tuning and a regional training objective that combines global flow matching, tumor‑region supervision, and stable foreground regularization. Ablation studies show that moderate tumor and stable‑foreground weighting improves the balance between image fidelity and tumor‑region accuracy, with the best configuration achieving a strong trade‑off using LoRA rank 64/64, MHA_max 25, λ_tumor 0.25, and λ_stable 0.1.
By Kamil Kwarciak, Marek Wodzinski
Loss-Based Active Learning for Neural Abstractive Summarization proposes LOBSTER, an active learning framework that selects unlabeled documents similar to the model’s high‑loss training examples to correct specific weaknesses. The method is tailored for abstractive summarization, addressing instability and computational bottlenecks seen in prior work. Experiments on three benchmark datasets and two backbone models show that LOBSTER matches or surpasses state‑of‑the‑art performance while speeding up query selection by up to 665×.
By Michail Ioannou, Tatiana Passali, George Michalopoulos, Grigorios Tsoumakas
ReFIne is a training framework that augments large reasoning models with three trustworthiness properties: interpretability, faithfulness, and reliability. It combines supervised fine‑tuning with GRPO to produce structured, tag‑based reasoning traces, explicitly disclose decisive information, and provide self‑assessments of soundness and confidence. Applied to Qwen3 models, ReFIne improves interpretability by 44.0 %, faithfulness by 18.8 %, and reliability by 42.4 % on mathematical benchmarks.
By Chung-En Sun, Ge Yan, Akshay Kulkarni, Tsui-Wei Weng
The paper investigates why vision‑language models like LLaVA‑1.5‑7B hallucinate objects in captions and proposes a targeted fix. By ranking attention heads whose image attention drops around hallucinated words, the authors identify 32 key heads and apply a head‑sliced LoRA adapter plus an inference‑time grounding controller. On COCO images, this combined method reduces hallucinated captions from 37% to 23% and hallucinated object mentions from 15.6% to 9.6%, while also lowering object recall.
By Armaan Sandhu, Abhilasha Senapati, Hima Kammachi