Hybrid language models combine attention with a fixed-size recurrent state, but the role of each channel remains unclear. We introduce two cache-level interventions. Split-prefill keeps only the KV cache or only the recurrent state from a prefilled context, then generates an answer. State-swap pairs the KV cache from one context with the recurrent state from another in a single forward pass. On Qwen3.5 and Falcon-H1, the two channels split cleanly. Exact retrieval survives only through attention (64–98% of full accuracy) and collapses to zero through recurrence. Output language and persona reverse the pattern: both survive recurrence (70–80%) while KV-only drops to 1% language accuracy. State-swap confirms this causally: the answer takes its value from the KV side and its language from the recurrent side. Recurrent-only generation also accepts words that were never in the context but share meaning or parts with seen items. Attention provides a lookup over what was said; the recurrent state shapes how the model says it next.

Fine-tuning LLMs introduces many important behaviors, such as instruction-following and safety alignment. This makes it crucial to study how fine-tuning changes models’ internal mechanisms. Sparse Autoencoders (SAEs) offer a powerful tool for interpreting neural networks by extracting concepts (features) represented in their activations. Previous work observed that SAEs trained on base models transfer effectively to instruction-tuned (chat) models, attributed to activation similarity. In this work, we propose feature drift as an alternative explanation: the feature space remains relevant, but the distribution of feature activations changes. In other words, fine-tuning recombines existing concepts rather than learning new ones. We validate this by showing base SAEs reconstruct both base and chat activations comparably despite systematic differences, with individual features exhibiting clear drift patterns. In a refusal behavior case study, we identify base SAE features that drift to activate on harmful instructions in chat models. Causal interventions using these features confirm that they mediate refusal. Our findings suggest that monitoring how existing features drift, rather than searching for entirely new features, may provide a more complete explanation of how fine-tuning changes model capabilities.

Process Reward Models (PRMs) have emerged as a promising approach for guiding large language models (LLMs) through multi-step reasoning by providing step-level feedback during inference. However, our evaluation across 7 LLMs reveals a failure mode: while PRMs improve performance for instruct mathematical models, they fail to enhance and sometimes degrade reasoning model performance. Through systematic analysis with linear probes, we identify distinct reward prediction patterns that differentiate reasoning from non-reasoning model outputs. To understand this mechanism, we train Sparse Autoencoders on the Qwen2.5-Math-PRM and analyze reasoning features. Our analysis reveals that 80% of these features respond to formatting artifacts (whitespace patterns, Unicode tokens, punctuation) rather than mathematical content. Reasoning model outputs exhibit distinct metacognitive patterns absent from standard mathematical solutions. This explains why they lead to unreliable reward estimation. Our findings expose a fundamental limitation in applying existing reward models to reasoning systems and provide mechanistic insights into this failure mode. We release our trained SAEs to facilitate future research into reward model interpretability.

Safety benchmarks for large language models (LLMs) commonly use multiple-choice question (MCQ) format. However, prior studies documented that MCQ format reduces refusal rates, which may significantly affect the safety behavior of the model. We provide a first mechanistic explanation of this effect on the Weapons of Mass Destruction Proxy (WMDP) dataset across two models – Llama-3.1-8B and Qwen3-8B. First, we use the tuned lens technique to show that when processing MCQ the model commits to an answer-letter token very quickly, which is in contrast to the gradual refusal build-up in open-ended formats. Additionally, we observe that activations on the MCQ format lie roughly 3× further from the refusal direction, compared with their open-ended counterparts. Moreover, ablating the attention heads that carry the MCQ format answer increases the refusal, which is a direct evidence that the MCQ answer-selection competes with refusal. These findings indicate that MCQ safety scores primarily measure knowledge capabilities and do not assess the safety-mechanisms.

The demand for efficient large language model inference has spurred interest in sparsification, yet current hardware support remains narrowly focused on 2:4 weight sparsity. In this work, we argue that activation sparsity despite being overlooked in hardware design offers a promising path for dynamic, input-adaptive compression with significant I/O and memory benefits. We present a comprehensive post-training study of N:M activation pruning across four LLMs (Llama2-7B-chat, Llama3.1-8B-Instruct, Qwen2.5-7B-Instruct, Gemma3-4B-Instruct), demonstrating that activation pruning consistently outperforms weight pruning at matched sparsity levels. We evaluate lightweight, plug-and-play error mitigation and selection strategies that require minimal or no calibration data across four sparsity patterns: 2:4, 4:8, 8:16, and 16:32. Among these, 16:32 approaches the performance of unstructured 50% sparsity and is is approximately 2.7× better than 2:4, while 8:16 offers an optimal balance of accuracy and practicality. Our results provide evidence that next-generation accelerators should consider native support for N:M activation sparsity and can serve as a strong baseline for the future methods.

Machine Unlearning (MU) is critical for removing private or hazardous information from deep learning models. While MU has advanced significantly in unimodal (text or vision) settings, multimodal unlearning (MMU) remains underexplored due to the lack of open benchmarks for evaluating cross-modal data removal. To address this gap, we introduce CLEAR, the first open-source benchmark designed specifically for MMU. CLEAR contains 200 fictitious individuals and 3,700 images linked with corresponding question-answer pairs, enabling a thorough evaluation across modalities. We conduct a comprehensive analysis of 11 MU methods (e.g., SCRUB, gradient ascent, DPO) across four evaluation sets, demonstrating that jointly unlearning both modalities outperforms single-modality approaches. The dataset is available at link

This paper presents a system developed for the Clinical NLP 2024 Shared Task, focusing on reliable text-to-SQL modeling on Electronic Health Records (EHRs). The goal is to create a model that accurately generates SQL queries for answerable questions while avoiding incorrect responses and handling unanswerable queries. Our approach comprises three main components: a query correspondence model, a text-to-SQL model, and an SQL verifier.For the query correspondence model, we trained a logistic regression model using hand-crafted features to distinguish between answerable and unanswerable queries. As for the text-to-SQL model, we utilized T5-3B as a pretrained language model, further fine-tuned on pairs of natural language questions and corresponding SQL queries. Finally, we applied the SQL verifier to inspect the resulting SQL queries.During the evaluation stage of the shared task, our system achieved an accuracy of 68.9 % (metric version without penalty), positioning it at the fifth-place ranking. While our approach did not surpass solutions based on large language models (LMMs) like ChatGPT, it demonstrates the promising potential of domain-specific specialized models that are more resource-efficient. The code is publicly available at https://github.com/runnerup96/EHRSQL-text2sql-solution.