The paper investigates how steering vectors influence large language models (LLMs) by conducting a mechanistic case study on refusal behavior. Using a multi-token activation patching framework, the authors find that steering methods primarily target the OV circuit of the attention mechanism, largely ignoring the QK circuit, and that these circuits are functionally interchangeable across different steering approaches. The study also shows that steering vectors can be sparsified by 85–96% with minimal performance loss and that key dimensions are consistently identified across methods.
By Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha
This paper introduces a multi‑branch fusion framework that combines transformer‑based semantics, rhetorical cues, stance representations, and psychologically motivated proxies to detect health misinformation and characterize its spread on online social networks. The authors propose an interpretable Cognitive Propagation Score (CPS) derived from text cues that estimate argument complexity, emotional intensity, and virality potential, aiding diffusion‑risk reasoning when engagement data are missing. Experiments on three benchmark datasets (Constraint, COVID‑19_FNIR, Monkeypox) demonstrate near‑perfect classification and ranking performance, with ablation studies showing complementary gains from psychological and rhetorical components.
By Mkululi Sikosana, Sean Maudsley-Barton, Oluwaseun Ajao
The paper investigates how well the implicit value rankings encoded by frontier AI systems—termed constitutional institutions—meet human demand. By auditing 23 large language model archetypes and surveying 1,649 U.S. participants, the authors find that user demand spans all five values (safety, helpfulness, honesty, autonomy, equity) but the supply is narrow, covering only about 2% of the demand space, with no model prioritizing helpfulness or autonomy. They propose a sparse two‑vertex menu that substantially reduces regret compared to the full set of models and formalize these observations as a budgeted‑pluralism trilemma.
whyItMatters":"The study reveals a significant mismatch between the values users prioritize and the values encoded by current AI models, highlighting the need for more diverse and aligned constitutional designs."
By Natalija Mitic, Soona Sedahmed A. O., Mamadou Selly Ly, Moustapha Cisse
FLaG (Frequency‑Domain Latent‑attention Gated Pooling) is a plug‑in token‑aggregation module that transforms encoder outputs into the Fourier domain, summarizes spectral tokens with learnable latent queries, applies a sample‑conditioned channel gate, and reconstructs modulated token representations for downstream pooling. The method is evaluated on antimicrobial peptide activity prediction, CIFAR‑10/100 image classification, and several RoBERTa language tasks, achieving state‑of‑the‑art performance on most metrics. Analyses show that FLaG emphasizes low‑frequency components while selectively amplifying high‑frequency signals in later layers, providing a transferable frequency‑domain bias across protein, visual, and textual representations.
By Kewei Li, Rongying Zhang, Xueli Wang, Xiwen Gong, Zhongjian Wang, Qiuchen Zhao, Lan Huang, Ruochi Zhang, Fengfeng Zhou
The paper investigates whether large language models (LLMs) can control their own internal representations, a question relevant to machine metacognition and AI safety. Previous work claimed such control using neurofeedback, but the targets were not privileged, allowing third‑party inference from prompts. By redesigning the paradigm to enforce privileged access—mirroring human neurofeedback experiments—the authors find that LLMs fail to reliably control these internal representations, indicating that earlier claims may have relied on superficial mechanisms.
By Koshiro Aoki, Ryota Takatsuki, Gouki Minegishi, Yusuke Haruki, Daisuke Kawahara
The paper introduces a framework that aligns self‑supervised respiratory encoders with medical terminology in a shared latent space, enabling zero‑shot respiratory sound classification. By using a medical LLM to generate structured reports from metadata, the method creates dense semantic anchors for contrastive learning, combining a sigmoid‑based contrastive loss with the encoder’s native SSL objective and similarity‑aware negative sampling. On nine tasks across six datasets, the approach achieves a 61.3% mean zero‑shot AUC, outperforming CLAP and Qwen2‑Audio, and reaches the highest linear probing AUC with only 43% of the data used by full‑scale baselines.
By Mustafa Talha \.Ilerisoy, Hung Manh Pham, Mathias Funk, Mykola Pechenizkiy, Aaqib Saeed
The paper investigates whether neuro‑symbolic (NeSy) AI models, which combine neural perception with symbolic reasoning, can mitigate backdoor attacks. It presents a systematic evaluation comparing the NeSy framework DeepProbLog to baseline neural networks across eight backdoor settings and four reasoning tasks. Results indicate that NeSy models are generally more robust than pure neural models, but their resilience depends heavily on how strictly the reasoning process is enforced and its alignment with the attack target.
By Marco Antonio Corallo, Andrea Agiollo, Mauro Conti, Alberto Giaretta
This paper investigates safety alignment in diffusion large language models (dLLMs), which generate text via iterative denoising instead of left‑to‑right decoding. By tracking token distributions and commitment decisions across denoising steps, the authors find that refusal signals are concentrated early in the denoising process and at leading response positions, and that early committed tokens strongly influence the final safety outcome. They introduce Refusal‑Aware Early Commitment (RAEC), a training‑free decoding method that preserves early refusal signals, and demonstrate that RAEC reduces attack success rates on LLaDA and Dream while largely maintaining utility.
By Guoli Wang, Haonan Shi, Tu Ouyang, An Wang
The paper introduces the MATCHA dataset, comprising 1,105 perceptual assessments from 83 experts on attribute-based music matches across five musical attributes—melody, harmony, rhythm, voice, and timbre. A triplet-based forced-choice experiment with 300 cases, including plagiarism, cover songs, and AI-generated music, was used to gather these judgments. Results show measurable agreement among participants and partial alignment with computational similarity measures, highlighting the need for perceptually grounded evaluation in generative AI for music.
By Roser Batlle-Roca, Woosung Choi, Joan Serr\`a, Fabio Morreale, Wei-Hsiang Liao, Xavier Serra, Emilia G\'omez, Yuki Mitsufuji
The paper investigates why reducing the modality gap between image and text representations in CLIP does not always improve zero‑shot classification accuracy. It shows that while average alignment improves, the relative decision margins among classes can shift, leading to a prediction‑level hubness where predictions concentrate on a few classes. Experiments across datasets confirm that accuracy drops correlate with increased prediction concentration for both linear and learning‑based gap corrections.
By Shota Sato, Hajime Kiyama, Tosho Hirasawa, Mamoru Komachi
FlexP-SFT introduces an aggregation-free framework for personalized split federated fine-tuning of large language models, eliminating the client-side aggregation step that traditionally causes communication bottlenecks and straggler issues. The method employs a layer‑flexible alignment strategy to balance personalization and generalization without global synchronization, and formulates split‑ratio selection as a resource‑aware discrete optimization problem. Experiments demonstrate that FlexP-SFT improves both accuracy and latency compared to baselines, achieving a superior resource‑accuracy trade‑off.
By Jiaxiang Geng, Tianjun Yuan, Pengchao Han, Ying Gao, Xianhao Chen, Bing Luo
BiasGym is a cost‑effective, generalizable framework that injects specific biases into large language models via token‑based fine‑tuning while keeping the model frozen. It then uses two debiasing methods—Scope and Steer—to identify and suppress or redirect the components responsible for biased behavior. The framework enables consistent bias elicitation, precise localization of bias associations, and targeted debiasing without harming downstream performance, and it has been shown to reduce real‑world stereotypes such as labeling Italians as reckless drivers.
By Sekh Mainul Islam, Nadav Borenstein, Siddhesh Milind Pawar, Haeun Yu, Arnav Arora, Isabelle Augenstein
The paper introduces patterning, a reweighting technique that adjusts preference pairs based on their susceptibility to bias, to debias a Gemma 2 9B Instruct reward model trained on Skywork-Reward-Preference v0.2. Using this method, the authors achieve a +14.2 ± 1.2 percentage point improvement on the RM‑Bench Hard split while maintaining overall accuracy, matching the best reported Hard‑split gain from a comparable model. The study also demonstrates that the learned weights are interpretable, transferable across Gemma variants, and partially effective on Llama 3.1 8B.
By George Wang, Elizabeth Donoway, Daniel Murfet
The paper identifies a problem in large language model (LLM) unlearning called forget‑set misalignment, where the set of data to be forgotten does not match what the model has actually memorized. Two failure modes are described: Under Unlearning, where memorized information is omitted from the forget set, and Out‑of‑Knowledge Unlearning, where the algorithm attempts to forget knowledge the model never learned, harming performance. The authors propose CONfs, a data‑blind framework that constructs model‑aligned forget sets by eliciting the model’s memorized knowledge, and demonstrate that it achieves near‑gold standard forgetting while preserving utility better than other data‑blind methods.
By Miso Kim, Georu Lee, Seungwon Jeong, Woojin Lee
The paper introduces EGT-KG, an evidence‑grounded typed knowledge graph retrieval framework designed to enhance scientific question answering with small language models (SLMs). It compares three QA settings—standard Retrieval‑Augmented Generation (RAG) and two EGT‑KG variants (automatically generated and expert‑defined relation schemas)—using a six‑dimensional evaluation on a biopolymer‑bound soil composite literature benchmark. Results show that both EGT‑KG variants outperform vanilla RAG, with the llama3:8b model achieving a final score of 70.37 (+14.67%) and 68.82 (+12.14%) for the AS and ES variants, respectively.
By Muran Yu, Jiechao Gao, Yuandong Pan, Barney H. Miao, Andrew C. Lesh, Kincho H. Law, Jie Wang, Michael D. Lepech
The paper examines how responsibility is assigned when AI systems fail, proposing a sociotechnical theory that distinguishes between AI incidents, organisational crises, and scandals. It argues that the configuration of an incident shapes actor-specific attribution, which in turn influences perceptions of capability, integrity, fairness, and relationships, and that public moralisation can elevate an incident to scandal. The authors introduce ‘accountable transparency’—a response framework combining timely notice, intelligible accounts, role acknowledgement, remedy, evidence of correction, and recourse—as a way to manage blame, trust, and communication credibility.
By Mohammad Saleh Torkestani, Taha Mansouri
The paper introduces a two‑stage training framework that combines Supervised Fine‑Tuning (SFT) and Direct Preference Optimization (DPO) to improve multimodal disaster severity assessment. It creates two datasets—ReasoningSet for validated rationales and PreferenceSet for paired rationales—using a single Human‑in‑the‑Loop workflow. Experiments on InternVL‑3‑8B and LLaVA‑1.5‑7B show that SFT boosts classification accuracy and Macro‑F1, while DPO further enhances interpretability and alignment with human judgment.
By Yuanjun Zhang, Fuzel Ahamed Shaik, Suvojit Acharjee, Fahad Khalid, Mourad Oussalah
The paper investigates how generative models for implied volatility surfaces must respect static no‑arbitrage constraints. By mapping latent codes to surfaces, it defines a scalar margin for each code and identifies the admissible latent set where this margin is nonnegative. The authors prove that admissible codes remain admissible under small perturbations, characterize the boundary via a zero‑margin condition, and derive a level‑set equation guiding dynamics toward this boundary. The framework applies broadly to models such as variational autoencoders and GANs, and numerical experiments confirm known boundaries in analytic examples and demonstrate how the boundary can correct violating latent codes.
By Jing Wang, Shuaiqiang Liu, Cornelis Vuik
FedReview is a review-based mechanism for federated learning that identifies and removes poisoned model updates without needing a server-side validation dataset or historic client knowledge. In each training round, the server randomly selects reviewers who evaluate incoming updates on their own training data, rank them, and estimate how many low-quality updates are likely poisoned. The server then aggregates these rankings via majority voting to filter out suspicious updates during model aggregation, enabling robust global model training even in adversarial settings.
By Tianhang Zheng, Yanlu Li, Bohan Deng, Baochun Li
The paper investigates why large language models lose safety alignment after benign fine‑tuning. It argues that safety alignment relies on a low‑rank, output‑routing geometry that becomes flatter during fine‑tuning, and that after only 100 benign examples this routing is sharpened in output‑side MLPs, leading to fragile safety while general performance remains relatively intact. Techniques like LoRA and ASAM can delay this collapse by reducing output‑side sharpness, but their effectiveness diminishes with larger fine‑tuning scales.
By Yitong Guo, Xiaoyi Chen, Siyuan Zhang, Xiaofeng Wang, Haixu Tang