{"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/ai-sandbagging-language-models-can","title":"AI Sandbagging: Language Models can Strategically Underperform on Evaluations","arxiv_id":"2406.07358","date":"2024-06-11","proceeding":null,"authors":["Teun van der Weij","Felix Hofstätter","Ollie Jaffe","Samuel F. Brown","Francis Rhys Ward"],"abstract":"Trustworthy capability evaluations are crucial for ensuring the safety of AI systems, and are becoming a key component of AI regulation. However, the developers of an AI system, or the AI system itself, may have incentives for evaluations to understate the AI's actual capability. These conflicting interests lead to the problem of sandbagging, which we define as strategic underperformance on an evaluation. In this paper we assess sandbagging capabilities in contemporary language models (LMs). We prompt frontier LMs, like GPT-4 and Claude 3 Opus, to selectively underperform on dangerous capability evaluations, while maintaining performance on general (harmless) capability evaluations. Moreover, we find that models can be fine-tuned, on a synthetic dataset, to hide specific capabilities unless given a password. This behaviour generalizes to high-quality, held-out benchmarks such as WMDP. In addition, we show that both frontier and smaller models can be prompted or password-locked to target specific scores on a capability evaluation. We have mediocre success in password-locking a model to mimic the answers a weaker model would give. Overall, our results suggest that capability evaluations are vulnerable to sandbagging. This vulnerability decreases the trustworthiness of evaluations, and thereby undermines important safety decisions regarding the development and deployment of advanced AI systems.","url_abs":"https://arxiv.org/abs/2406.07358v4","url_pdf":"https://arxiv.org/pdf/2406.07358v4.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":"ai-sandbagging-language-models-can","repo_url":"https://github.com/teunvdweij/sandbagging","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[],"methods":[{"method_slug":"absolute-position-encodings","method_name":"Absolute Position Encodings"},{"method_slug":"adam","method_name":"Adam"},{"method_slug":"attention","method_name":"Attention"},{"method_slug":"bpe","method_name":"BPE"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"dropout","method_name":"Dropout"},{"method_slug":"gpt-4","method_name":"GPT-4"},{"method_slug":"label-smoothing","method_name":"Label Smoothing"},{"method_slug":"layer-normalization","method_name":"Layer Normalization"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"multi-head-attention","method_name":"Multi-Head Attention"},{"method_slug":"position-wise-feed-forward-layer","method_name":"Position-Wise Feed-Forward Layer"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"transformer","method_name":"Transformer"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2406.07358","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"2406.07358"}},"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/teunvdweij/sandbagging","reach":{"status":"ok"}}],"summary":{"ran":5},"by_repo_kind":{"official":{"samples":5,"ran":5,"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":5,"samples":[{"code_sha256_prefix":"4130e9405a2f32c9","entry":"count_correct_answers","repo":"teunvdweij/sandbagging","repo_kind":"official","path":"plotting/plot_weaker_model_emulation_results.py","file_url":"https://github.com/teunvdweij/sandbagging/blob/HEAD/plotting/plot_weaker_model_emulation_results.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"4130e9405a2f32c9"}},{"code_sha256_prefix":"a08bd3d753162520","entry":"get_base_mc_prompt","repo":"teunvdweij/sandbagging","repo_kind":"official","path":"src/wmdp_sandbagging/get_benchmark_results_for_hf_models.py","file_url":"https://github.com/teunvdweij/sandbagging/blob/HEAD/src/wmdp_sandbagging/get_benchmark_results_for_hf_models.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"a08bd3d753162520"}},{"code_sha256_prefix":"e140939aab3ae1c9","entry":"get_sequence_bias","repo":"teunvdweij/sandbagging","repo_kind":"official","path":"src/wmdp_sandbagging/get_benchmark_results_for_hf_models.py","file_url":"https://github.com/teunvdweij/sandbagging/blob/HEAD/src/wmdp_sandbagging/get_benchmark_results_for_hf_models.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"e140939aab3ae1c9"}},{"code_sha256_prefix":"d24e1905ab5edbc9","entry":"init_tokenizer","repo":"teunvdweij/sandbagging","repo_kind":"official","path":"old/model.py","file_url":"https://github.com/teunvdweij/sandbagging/blob/HEAD/old/model.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"d24e1905ab5edbc9"}},{"code_sha256_prefix":"d332f3065458bbd9","entry":"load_sequence_from_jsonl","repo":"teunvdweij/sandbagging","repo_kind":"official","path":"src/utils.py","file_url":"https://github.com/teunvdweij/sandbagging/blob/HEAD/src/utils.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"d332f3065458bbd9"}}]},"arxiv_metadata":null,"syntology_extracted_results":null}