This work uses answer logits at the end of each reasoning step to estimate each step's causal importance to the final answer and intermediate guesses, shedding light on the answer formation process of several reasoning model families.
Daniel Scalena, Sara Candussio, Luca Bortolussi et al.· arXiv.org· 3 citations
It is demonstrated that patching a subset of a few CoT tokens with their original activations recovers near-perfect full-trace performance, providing causal evidence that task information is not concentrated in a small set of CoT tokens identifiable by heuristics, but rather distributed across the full reasoning chain.
Sara Candussio, Daniel Scalena, Luca Bortolussi et al.· arXiv.org· 0 citations
It is argued that this question cannot be answered if this question cannot be answered if a clearer distinction between language models designed as technical products and language models designed as tools for studying language itself is considered.
It is suggested that biography leakage into generated turns is an artefact of how interlocutor visibility is configured across training and inference, and separating the three stages is necessary.
Daniela Occhipinti, Malvina Nissim, M. Guerini· 0 citations
CALAMITA is conceived as a rolling benchmark, enabling continuous integration of new tasks and models, and argues that this combination offers a blueprint for other languages and communities seeking inclusive and rigorous LLM evaluation practices.
Malvina Nissim, Danilo Croce, V. Patti et al.· Italian Journal of Computati...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
