{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/robustly-identifying-concepts-introduced","title":"Robustly identifying concepts introduced during chat fine-tuning using crosscoders","arxiv_id":"2504.02922","date":"2025-04-03","proceeding":null,"authors":["Julian Minder","Clement Dumas","Caden Juang","Bilal Chugtai","Neel Nanda"],"abstract":"Model diffing is the study of how fine-tuning changes a model's representations and internal algorithms. Many behaviours of interest are introduced during fine-tuning, and model diffing offers a promising lens to interpret such behaviors. Crosscoders are a recent model diffing method that learns a shared dictionary of interpretable concepts represented as latent directions in both the base and fine-tuned models, allowing us to track how concepts shift or emerge during fine-tuning. Notably, prior work has observed concepts with no direction in the base model, and it was hypothesized that these model-specific latents were concepts introduced during fine-tuning. However, we identify two issues which stem from the crosscoders L1 training loss that can misattribute concepts as unique to the fine-tuned model, when they really exist in both models. We develop Latent Scaling to flag these issues by more accurately measuring each latent's presence across models. In experiments comparing Gemma 2 2B base and chat models, we observe that the standard crosscoder suffers heavily from these issues. Building on these insights, we train a crosscoder with BatchTopK loss and show that it substantially mitigates these issues, finding more genuinely chat-specific and highly interpretable concepts. We recommend practitioners adopt similar techniques. Using the BatchTopK crosscoder, we successfully identify a set of genuinely chat-specific latents that are both interpretable and causally effective, representing concepts such as $\\textit{false information}$ and $\\textit{personal question}$, along with multiple refusal-related latents that show nuanced preferences for different refusal triggers. Overall, our work advances best practices for the crosscoder-based methodology for model diffing and demonstrates that it can provide concrete insights into how chat tuning modifies language model behavior.","url_abs":"https://arxiv.org/abs/2504.02922v1","url_pdf":"https://arxiv.org/pdf/2504.02922v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"robustly-identifying-concepts-introduced","repo_url":"https://github.com/jkminder/dictionary_learning","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[],"methods":[{"method_slug":"adopt","method_name":"ADOPT"},{"method_slug":"base","method_name":"BASE"},{"method_slug":"set","method_name":"SET"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2504.02922","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"2504.02922"}},"developers":"https://syntology.ai/developers","read_at":"2026-09-24T18:15:14+00:00","read_at_is":"when the build read Syntology's graph, not when any sample ran","claim":"Per-sample execution status on synthesized fixtures; not a correctness claim about the paper. Samples come from repositories linked to the paper, official or community; repo_kind says which.","repos":[{"provenance":"external:paperswithcode_snapshot_2025-07-28","url":"https://github.com/jkminder/dictionary_learning","reach":{"status":"ok","spdx":"MIT"}}],"summary":{"unverified":8},"by_repo_kind":{"official":{"samples":8,"ran":0,"repositories":1}},"repo_kind_vocabulary":{"official":"The archive marks this repository official for the paper","named_in_paper":"The archive records that the paper mentions this repository; it is not marked official","listed":"In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper","found_in_text":"Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted","community":"Not in the archive's code links for this paper; a community repository Syntology harvested"},"n_pointer_only_for_licence":0,"samples":[{"code_sha256_prefix":"47523c58bac9ebdc","entry":"feature_effect","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/interp.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/interp.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"47523c58bac9ebdc"}},{"code_sha256_prefix":"b9097b889cb4d77e","entry":"get_model","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/training.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/training.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"b9097b889cb4d77e"}},{"code_sha256_prefix":"87467b7150f616a8","entry":"get_stats","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/training.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/training.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"87467b7150f616a8"}},{"code_sha256_prefix":"5712034dc298886f","entry":"grad_pursuit","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/grad_pursuit.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/grad_pursuit.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"5712034dc298886f"}},{"code_sha256_prefix":"624c94649d7e86d4","entry":"hf_dataset_to_generator","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/utils.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/utils.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"624c94649d7e86d4"}},{"code_sha256_prefix":"77a35fa07b4643b5","entry":"numpy_to_torch_dtype","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/utils.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/utils.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"77a35fa07b4643b5"}},{"code_sha256_prefix":"2f3727370354fa2b","entry":"run_validation","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/training.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/training.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"2f3727370354fa2b"}},{"code_sha256_prefix":"c3489b6eb423ae43","entry":"zst_to_generator","repo":"jkminder/dictionary_learning","repo_kind":"official","path":"dictionary_learning/utils.py","file_url":"https://github.com/jkminder/dictionary_learning/blob/HEAD/dictionary_learning/utils.py","link_basis":"harvester_set","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"MIT","inline_ok":true,"mcp_get_code":{"code_sha256":"c3489b6eb423ae43"}}]},"arxiv_metadata":null,"syntology_extracted_results":null}