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.

Recent progress in language models has been defined by scale, yet many practical applications benefit more from robust reasoning than extensive parametric knowledge. We introduce OCC-RAG, a family of models for faithful context-based QA, which prioritizes task-specific reasoning capabilities over knowledge capacity. To train OCC-RAG, we implement a novel pipeline synthesizing over three million multi-context, multi-hop QA examples that target multi-hop reasoning, strict context faithfulness, and calibrated abstention. Mid-trained on this corpus, the models produce structured reasoning traces with source citations grounded in literal quotes from the context. Despite their compact size, OCC-RAG models outperform their same-size general-purpose counterparts on every benchmark and exceed models 2–6× larger on financial reasoning (TAT-QA) and faithfulness (ConFiQA), attaining the best results among all evaluated open-weight models up to 32B parameters. On ConFiQA they rank behind only a frontier proprietary models; on multi-hop reasoning (HotpotQA, MuSiQue) and refusal (MuSiQue-Un) they are on par with 8B-class models and close much of the remaining gap to the largest evaluated systems. We release both our code and family of models.[<https://hf.co/collections/occ-ai/occ-rag>]

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.