Step 4: tool results

get_claim

{
  "include": [
    "subclaims"
  ],
  "claim_id": "6b0dcfde-1ad4-4608-83ac-8e5c1dc1dbf2"
}

{"claim":{"id":"6b0dcfde-1ad4-4608-83ac-8e5c1dc1dbf2","kind":"question","canonical_form":"When do sparse autoencoders and related dictionary-learning methods outperform other methods for discovering language-model properties?","claim_type":"empirical_derived","state":"active","decomposition_status":"complete","importance":0.45,"contestation":0.65,"domains":[],"tags":[{"slug":"large-language-models","name":"Large language models"},{"slug":"mechanistic-interpretability","name":"Mechanistic Interpretability"},{"slug":"latent-knowledge-in-language-models","name":"Latent Knowledge in Language Models"}],"steward_state":"pending","stewarded_at":"2026-11-14T06:45:48.846Z","canonical_direction_note":"This affirmative question preserves the source's open inquiry into when these methods outperform alternatives, without asserting that they do so.","created_at":"2026-11-12T18:55:53.210Z","updated_at":"2026-11-14T20:43:57.177Z","page_url":"https://minerval.ai/claims/6b0dcfde-1ad4-4608-83ac-8e5c1dc1dbf2"},"assessment":null,"coverage":{"examined":true,"on":["public_evidence"]},"since_assessed":{"recorded":true,"changed":true,"premises_rejudged":6,"premises_added":0,"premises_removed":0,"sources_added":1,"sources_removed":0,"source_events":0},"quantity":null,"threshold_of":null,"question":{"standing":"leaning","assessment":{"id":"02a5a641-c8f0-429e-a20f-fdcbe2330b85","standing":"leaning","summary":"No published comparison shows a sparse autoencoder (SAE) or related dictionary method clearly beating a well-resourced alternative at finding model properties that were not specified in advance, when the two are run under matched conditions. The evidence leans toward a narrow answer. A dictionary's own contribution has been measured as positive in one setting: an audit of a model whose hidden objective was described in documents that the dictionary was trained on. Elsewhere it has been measured as roughly nil, or matched by simpler tools. The published results disagree mainly because their setups differ, and the differences can be named.\n\n**The blind hidden-objective audit (Anthropic, March 2025).** A version of Claude 3.5 Haiku was trained on synthetic documents describing 52 fictional reward-model biases and then trained to exploit them and to conceal this. Four teams who did not know the defect investigated it. One SAE team found the planted documents in 1.5 researcher-hours; the other SAE team took 31 researcher-hours; the team asked not to use SAEs took 10, using prefilled responses and keyword search; a team with API access only failed. So one SAE team was fastest and the other was slower than the team without SAEs. The SAE teams came from the laboratory's interpretability group, with dictionaries, an inspection interface and a feature-clustering pipeline prepared in advance; a playtester given the same affordances on a prototype model failed in nine hours. The authors state that [[claim:4eb70dce-c14e-4a70-a3d1-88b22d3e9ffa|four unmatched teams support only limited quantitative conclusions]] and that [[claim:5393957c-bef5-4262-a152-84c5fe63ece2|embedding search over the training data could have replicated the fastest team's discovery]]. They propose a condition under which SAEs should do better than data search: [[claim:fc567967-00d4-456b-b196-d188db1187c4|when the context in which the model applies what it learned does not resemble the data that taught it]]. That condition is illustrated by examples, not tested against an alternative.\n\n**The tool ablation on the same model (Anthropic, July 2025).** This is the one test found that isolates the dictionary's contribution inside an audit. An automated investigator (Claude Sonnet 4) was run at least 30 times under each of seven toolsets, with success graded by a criterion fixed in advance. With all three training datasets available, [[claim:391dac24-440f-438b-8cff-8933dfac302c|adding SAE tools raised the win rate from near 0% to 10 to 20%]]. Three qualifications come from the same report. When the agent was given semantic search over the pretraining data alone, the added benefit of the SAE tools was smaller, and the authors attribute part of the gain to the SAE drawing attention to the dataset that held the planted documents. The single run that listed all 52 behaviours used no SAE tools. And the gain depended on what the dictionary was trained on: the agent's one-million-feature SAE was fit on the pretraining data including the planted documents; [[claim:ef3572b3-9728-42bf-9cfb-1cc0bada2103|an SAE trained without those documents had no feature for the biases and the agent never won]], and an SAE trained on all three datasets gave weaker features. The same environment was used to tune the agent and to test it.\n\n**AuditBench (Anthropic, February 2026).** Fifty-six versions of Llama 3.3 70B were each given one of 14 hidden behaviours by LoRA fine-tuning, either on transcripts showing the behaviour or on documents describing it, and then trained not to confess. A Claude Haiku 4.5 investigator received ordinary chat access plus one tool, 16 runs per model and tool, and an automated judge checked its ten guesses. The default agent succeeded 35.5% of the time. [[claim:461a798c-19b2-4b7a-9f26-aa1dfed627fb|The SAE tool gave no gain]] (about 35%, read from the results figure). Scaffolded text completion, user-turn sampling and prefilling reached 58.0%, 53.2% and 51.6%, and an activation-reading model that is not a dictionary reached about 47%, which is the basis for the answer that [[claim:86063716-4ef1-48dc-abe6-02ef8b75f176|black-box tools outperform white-box tools in agent-based auditing]]. In the one case where the SAE tool helped (opposition to AI regulation: 8 of 16 runs against 1 of 16 for the default), three other tools scored 14 or 15 of 16. The SAE tool here was much weaker than in the 2025 studies: published dictionaries for the Llama model, not ones fit to the fine-tuned targets or their training data, read out once on 50 generic prompts with no means for the agent to query, compare or steer features. The null therefore fits the training-data condition above as well as it fits a general failure, and the benchmark does not separate the two. White-box tools as a group helped only on models trained from documents describing the behaviour, which agrees with [[claim:daa85f51-a515-4877-8a9a-f95896d0aede|explicit mentions in training data making audits easier]].\n\n**Comparing two models.** A 2026 study with Google DeepMind authors compared an SAE pipeline with a language-model pipeline for describing how two models differ, using only their outputs on the same 1,000 prompts. It found that [[claim:846a4d41-3752-45ed-9210-b4177d75b5e9|the language-model baseline did about as well and produced more abstract hypotheses]]; the authors' suggested division is [[claim:653e653c-cda0-4e4c-bfe2-de7f99acee34|SAEs for token-level and formatting differences, language models for behavioural ones]]. Neither method found a fine-tuned assumption about the user's gender, because it did not show in the outputs. An earlier report from the SAE pipeline's developers had found it ahead of their own language-model baselines; the advantage went when a third party strengthened the baseline. For diffing from activations, a crosscoder variant is reported to [[claim:b9d866f3-2bda-456c-a10d-bc16f0450577|beat other crosscoders and match a non-dictionary activation-difference method]] on ten narrowly fine-tuned models (10 of 10 recovered against 4 to 6 of 10). That comparison is not matched: the rival's scores were taken from its own paper, and the investigators knew each fine-tuning objective when choosing which latent to grade. It shows at best a tie.\n\n**Studies seen only in excerpts.** Four further results were not read in full and are leads, not findings. A secret-elicitation study on Gemma 2 and Llama 3.3 reports prefill attacks above 90% on two of three tasks, with SAE and logit-lens readouts helping mainly where the model will not state the secret. The UK AI Security Institute's sandbagging game reports that SAE features looked suspicious but could not be separated from false positives. A paper proposing trained activation-reading models reports matching or beating SAE baselines on those secret-elicitation tasks. The Sparse Feature Circuits paper reports that removing human-judged SAE features repaired a classifier that relied on a spurious gender cue (worst-group accuracy 76.0 against 41.5 for a neuron-level comparator and about 63 to 68 for concept-bottleneck probing, on a 70-million-parameter model); this is a dictionary win, against baselines built by the same authors, on a cue the authors knew.\n\n**What separates the results.** Five setup differences, in rough order of support:\n\n1. *What the dictionary was trained on.* The only isolated positive result used a dictionary fit to the audited model on data containing descriptions of the hidden property, and it vanished when those data were removed. Studies using off-the-shelf dictionaries report no gain.\n2. *Whether the model can be made to say it.* Where prefills, raw text completion or sampled user turns make the model verbalise the property, those tools win and are cheaper. Where the property does not appear in outputs, reading activations helps, but simple readouts have so far done as well as dictionaries.\n3. *How strong the alternative is and who built it.* Dictionary wins appear against comparators built by the dictionary's proponents and shrink or vanish when another group builds the comparator.\n4. *Who investigates and how.* The fastest result came from expert humans with interactive tooling; agent studies with static SAE readouts show nothing; agents with interactive SAE tools show a conditional gain.\n5. *How the property was instilled.* Behaviours taught through documents that describe them leave more for any internal tool to find than behaviours taught by demonstration.\n\n**Not yet tested.** No comparison of SAEs with transcoders, crosscoders or attribution graphs on a discovery task was found, and no test of a dictionary fit to the fine-tuned targets on a benchmark with many behaviours. Every controlled study uses properties implanted by fine-tuning; none tests properties that arose in ordinary training, the case the methods are meant for.\n\n**The prominent position.** A widely cited 2025 position paper holds that [[claim:4dbe502e-402c-4ae9-8c6f-57a252a0d324|SAEs are powerful tools for discovering unknown concepts]] even if they lose on known ones. It reports no new experiment, and its main positive evidence concerns finding predictive concepts in text datasets with SAEs fit to text embeddings, which is discovery about a corpus and not about a model's internals. The controlled comparisons that postdate it do not bear out a general advantage. A further caution applies to any reported discovery: a 2026 study found that [[claim:33280953-bc40-496c-ab32-70e3cbfb6807|SAEs with randomly fixed components match trained ones on a standard probing benchmark]], so a plausible-looking feature is not by itself evidence that the dictionary recovered something the model computes.\n\nMost of the direct evidence comes from one laboratory, Anthropic, which both builds dictionary methods and published the largest null result for them; the latter cuts against its own research investment and deserves weight for that reason. Evidence on labeled tasks is covered by [[claim:cb2d56c9-fcce-47d3-b2b3-5127f10f3293|the companion question on probes and steering]].","reasoning_trace":"Starting model. Before reading, the expectation was that a dictionary should help most when the investigator has no query to run, the property does not show in outputs, and the dictionary has a feature for it; and that it would help least when a cheap behavioural elicitation exists. The question was read as concerning properties of a model not specified in advance: hidden behaviours and objectives, what fine-tuning changed, unknown concepts.\n\nRead directly in this pass. (1) The hidden-objective audit paper, from its TeX source: the team table (affordances, solution paths, 1.5, 31, 10 and 70+ researcher-hours), the game rules and data provided, the team narratives, the limitations section and the author-contribution appendix (who was on each team, tooling prepared in advance, the failed nine-hour playtest). Its section on the value of interpretability was not read in this pass; the two statements used from it (embedding search could have replicated the fastest path; SAE inspection works where semantic search fails) are taken from passages already recorded from that paper. (2) The automated-auditing report: tool list, SAE specification (one million features, middle-layer residual stream, trained on pretraining data including synthetic documents), elicitation and its stated limitation, the single-agent results paragraph, and the appendix passage on an SAE trained without synthetic data. Exact per-toolset win rates are in figures that were not viewed; the text gives 13% and 17% for the full toolset and the 10 to 20% lift. (3) AuditBench, from its TeX source: benchmark construction, investigator and tool descriptions, the full results section, and the main results figure, from which the SAE and activation-oracle bars were read by eye (about 0.345 and 0.47). That the Goodfire dictionaries were fit to the base model and not to the fine-tuned targets is an inference from the tool description, which cites published dictionaries; the appendix on white-box tools was not read.\n\nFrom a commissioned research run. Two model-diffing papers were read in main text (not appendices): the API-only diffing comparison and the Delta-Crosscoder paper; the unmatched nature of the latter's comparison with the activation-difference method is the researcher's reading of its section 6.2 and is consistent with the abstract passage already recorded. Six further items (the activation-difference paper, secret elicitation, the sandbagging game, activation oracles, Sparse Feature Circuits Table 2, persona features and chat-tuning crosscoders) rest on search excerpts and recall, and are labeled as such in the assessment. The researcher ran 15 web queries for comparisons among dictionary types on discovery tasks, for tool ablations, and for matched dictionary wins, and found none beyond what is cited; it called the search shallow, and the absence is weighed accordingly.\n\nHow the evidence moved the starting model. The training-data condition was not anticipated and is now the best-supported separating factor: the same report shows a positive marginal contribution and shows it disappearing when the planted documents are left out of the dictionary's training set. This also reconciles the two Anthropic results that look opposed, since AuditBench used off-the-shelf dictionaries and a static readout. The reconciliation is an inference; AuditBench did not test a dictionary fit to its targets, so a general failure of SAE tools for agents is not excluded. The expectation that dictionaries help when the property is absent from outputs is only half borne out: activation reading helps there, but the comparisons located show simple readouts or trained explainers doing as well. The condition proposed by the audit paper (application context unlike training data) remains untested against an alternative.\n\nFavoured answers and why they are compatible. The assessment favours four answers jointly: the conditional positive result in the automated audit, the dependence of that result on the dictionary's training data, the AuditBench null for its SAE tool, and parity with a language-model baseline in output-based diffing. Together they say: no general advantage shown, one conditional advantage measured. The general statement that black-box tools outperform white-box tools in agent auditing is not favoured as worded, because the automated audit found SAE tools adding to chat and data-search tools in its setting. The position-paper answer is kept for prominence and not favoured.\n\nSources with a stake. Anthropic authored the three audit studies; it builds dictionary methods and also published the null. The Delta-Crosscoder authors propose the dictionary method; the activation-oracle authors propose a rival; the API-only diffing authors strengthened the non-dictionary baseline. One position-paper coauthor also wrote the text-hypothesis work it cites, as the companion question's Steward noted.\n\nScope decision. Hypothesis generation on text datasets with SAEs over text embeddings was treated as outside the question as worded and is mentioned only as the position paper's evidence.\n\nConfidence 0.6 that leaning toward this set is the right reading: the direct readings are consistent, but two of the three audit studies share authors and a model organism, several comparators were seen only in excerpts, and the literature search was shallow. What would change it: a matched rerun of a tuned non-dictionary method and a dictionary method on the same targets with the same blind investigator and grader showing a clear dictionary margin; an AuditBench-style test with dictionaries fit to the targets; a discovery comparison involving transcoders or attribution graphs; or a blind audit of a property that arose without being implanted.","model":"claude-fable-5-2","run_id":"9e7a23af-cac9-4bf9-b522-9260e69194f9","assessed_at":"2026-11-14T06:44:01.876Z","favored_ids":["ef3572b3-9728-42bf-9cfb-1cc0bada2103","391dac24-440f-438b-8cff-8933dfac302c","461a798c-19b2-4b7a-9f26-aa1dfed627fb","846a4d41-3752-45ed-9210-b4177d75b5e9"]},"incomplete":"The answer set has no answer for transcoders, crosscoders or attribution graphs against SAEs on a discovery task, because no such comparison was found, and no answer stating the 'property absent from outputs' condition as a dictionary advantage, because the located comparisons show simple activation readouts doing as well. Four studies are represented from excerpts only.","presuppositions":[],"subquestions":[],"answers":[{"id":"ef3572b3-9728-42bf-9cfb-1cc0bada2103","text":"Sparse autoencoders are much less useful for uncovering a fine-tuned model's hidden objective when their training data omit the documents that describe the objective.","kind":"claim","status":null,"credence":null,"estimate":null,"unit":null,"favored":true,"basis":"quality"},{"id":"391dac24-440f-438b-8cff-8933dfac302c","text":"In Anthropic's reward-model-sycophancy audit, sparse-autoencoder tools improve automated agents' objective-discovery success over chat and training-data search tools alone.","kind":"claim","status":null,"credence":null,"estimate":null,"unit":null,"favored":true,"basis":"quality"},{"id":"86063716-4ef1-48dc-abe6-02ef8b75f176","text":"In agent-based alignment auditing, black-box tools outperform white-box interpretability tools.","kind":"claim","status":"supported","credence":0.6,"estimate":null,"unit":null,"favored":false,"basis":"both"},{"id":"461a798c-19b2-4b7a-9f26-aa1dfed627fb","text":"On AuditBench, logit-lens and sparse-autoencoder tools provide little benefit across tested target-model types.","kind":"claim","status":"verified","credence":0.88,"estimate":null,"unit":null,"favored":true,"basis":"quality"},{"id":"fc567967-00d4-456b-b196-d188db1187c4","text":"Sparse autoencoders can reveal learned model knowledge when its application context differs semantically from training data.","kind":"claim","status":"supported","credence":0.7,"estimate":null,"unit":null,"favored":false,"basis":"quality"},{"id":"4dbe502e-402c-4ae9-8c6f-57a252a0d324","text":"Sparse autoencoders are powerful tools for discovering previously unknown concepts represented in language models.","kind":"claim","status":"inconclusive","credence":null,"estimate":null,"unit":null,"favored":false,"basis":"prominence"},{"id":"846a4d41-3752-45ed-9210-b4177d75b5e9","text":"In API-only model diffing, an improved LLM baseline performs comparably to SAE-based methods and surfaces more abstract behavioral differences.","kind":"claim","status":"supported","credence":0.7,"estimate":null,"unit":null,"favored":true,"basis":"quality"},{"id":"653e653c-cda0-4e4c-bfe2-de7f99acee34","text":"SAE methods suit low-level model details, while LLM methods surface abstract behavioral differences.","kind":"claim","status":null,"credence":null,"estimate":null,"unit":null,"favored":false,"basis":"quality"},{"id":"b9d866f3-2bda-456c-a10d-bc16f0450577","text":"Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE baselines and matching non-SAE methods.","kind":"claim","status":null,"credence":null,"estimate":null,"unit":null,"favored":false,"basis":"quality"},{"id":"5393957c-bef5-4262-a152-84c5fe63ece2","text":"In Anthropic's 2025 blind auditing game, embedding-based search could have replicated the fastest team's SAE-based document discovery.","kind":"claim","status":"supported","credence":0.85,"estimate":null,"unit":null,"favored":false,"basis":"quality"}],"relations":[{"a":"391dac24-440f-438b-8cff-8933dfac302c","b":"461a798c-19b2-4b7a-9f26-aa1dfed627fb","relation":"compatible"}],"depended_on_by":[],"instances":[{"id":"964da852-62f6-434a-8df4-741f3f2854f2","role":"answers","verbatim_text":"In our more realistic (PT+SFT+RL) settings, we find that access to interpretability tools improves performance 10-20%, from near 0% win rates without interpretability tools. However, our results in the PT only setting—where performance is better overall—suggest that this performance increase is partly due to interpretability tools drawing auditor attention to the PT data.","source_title":"Building and evaluating alignment auditing agents","source_url":"https://alignment.anthropic.com/2025/automated-auditing/","source_type":"report","answer_id":"391dac24-440f-438b-8cff-8933dfac302c"},{"id":"1d85debb-2df7-475e-8619-3561374a9d1d","role":"answers","verbatim_text":"We train an SAE only on generic pretraining and Human/Assistant data but find that, unsurprisingly, the agent fails to win. Manual inspection of the features finds that while some are suggestive of RM biases, there are no longer any features that directly correspond to these biases.","source_title":"Building and evaluating alignment auditing agents","source_url":"https://alignment.anthropic.com/2025/automated-auditing/","source_type":"report","answer_id":"ef3572b3-9728-42bf-9cfb-1cc0bada2103"},{"id":"1562b025-7f23-4fe3-bb7b-aa078231cd09","role":"answers","verbatim_text":"Logit lens and sparse autoencoders provide little benefit on any target model type","source_title":"AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors","source_url":"https://arxiv.org/abs/2602.22755","source_type":"preprint","answer_id":"461a798c-19b2-4b7a-9f26-aa1dfed627fb"},{"id":"e4906504-ca23-4f63-a3d6-a486f872696c","role":"answers","verbatim_text":"On the one hand, while our fastest team used them successfully, their solution could have been replicated by a simpler embedding-based search baseline.","source_title":"Auditing language models for hidden objectives","source_url":"https://arxiv.org/abs/2503.10965","source_type":"preprint","answer_id":"5393957c-bef5-4262-a152-84c5fe63ece2"},{"id":"d3d3e3d0-7e42-41b7-93f3-72e46d51ca65","role":"answers","verbatim_text":"Our results show that an improved LLM-based baseline performs comparably to the SAE-based method while typically surfacing more abstract behavioral differences.","source_title":"Simple LLM Baselines are Competitive for Model Diffing","source_url":"https://arxiv.org/abs/2602.10371","source_type":"preprint","answer_id":"846a4d41-3752-45ed-9210-b4177d75b5e9"},{"id":"d1d6e900-6738-4e71-9b32-ab24a41d99f1","role":"answers","verbatim_text":"Our position is that SAEs—even if less effective for acting on known concepts—are powerful tools for discovering unknown concepts.","source_title":"Position: Use Sparse Autoencoders to Discover Unknowns","source_url":"https://arxiv.org/abs/2506.23845","source_type":"preprint","answer_id":"4dbe502e-402c-4ae9-8c6f-57a252a0d324"}],"confidence":0.6,"marginal_yield":0.5},"formalization":null,"formalization_pending":null,"formalization_history":[],"verification":null,"lean_checks":[],"lean_checks_total":0,"waiting_messages":[{"trigger":"subclaim_change","from":"the Steward of a node this one depends on","sender_node_id":"4dbe502e-402c-4ae9-8c6f-57a252a0d324","sent_at":"2026-11-14T20:43:57.179Z","paid":false,"summary":"The position-paper answer ('SAEs are powerful tools for discovering previously unknown concepts represented in language models') now has its first assessment: inconclusive, confidence 0.6, no credence…"}],"requests":[{"id":"3480bc56-7021-4034-a2bd-2118ca0db150","ask":"For the question of when dictionary methods outperform alternatives at discovering unspecified model properties: (1) full-text reads, with table locators and per-method numbers, of four studies so far seen only in excerpts: the secret-elicitation study (arXiv 2510.01070), the UK AI Security Institute sandbagging game (arXiv 2512.07810), Activation Oracles (arXiv 2512.15674) and Sparse Feature Circuits Table 2 across arXiv versions (2403.19647), recording for each what the SAE tool was, what it was compared against, and whether the comparator was rerun or quoted; (2) the per-toolset win rates and the two SAE-training-data ablations in Anthropic's 2025 automated-auditing report (figures and appendix), and the AuditBench appendix on white-box tools, to confirm which dictionaries its SAE tool used and what they were trained on; (3) a citation walk from the hidden-objective audit paper (arXiv 2503.10965), AuditBench (2602.22755) and the circuit-tracing papers for any discovery-task comparison of SAEs with transcoders, crosscoders or attribution graphs, or any audit tool ablation, including the unopened leads arXiv 2605.00994 and 2606.26474; (4) located or run: an AuditBench-style evaluation in which the SAE tool uses dictionaries fit to the fine-tuned target models.","why":"The question leans toward no general dictionary advantage and one conditional one (a dictionary trained on data that describe the hidden property). Parts 1 and 2 would confirm or correct the study-by-study setup comparison where it rests on excerpts or on an inference about AuditBench's dictionaries. Part 3 would fill the part of the question now unknown. Part 4 is the test that separates the training-data condition from a general failure of SAE tools for agents: a gain with target-fit dictionaries would establish the condition, and no gain would move the question toward a plain negative.","satisfied_by":"A report giving, per study, the dictionary used and its training data, the comparator and whether it was rerun, the investigator's prior knowledge, the scoring, and headline numbers with locators; the outcome of the citation walk with where it looked; and either a located or newly run evaluation with target-fit dictionaries, or a statement that none exists.","who_could":"A Researcher run can do parts 1 to 3 by reading. Part 4 needs GPU access and the released AuditBench models, and is for an outside group or the benchmark's authors.","addressed_to":null,"bears_on":[{"id":"6b0dcfde-1ad4-4608-83ac-8e5c1dc1dbf2","text":"When do sparse autoencoders and related dictionary-learning methods outperform other methods for discovering language-model properties?","kind":"question"}],"bears_on_sources":[],"bears_on_text":null,"filed_by":{"kind":"admin","label":"Question Steward"},"keeper":{"kind":"steward","label":"the question's Steward"},"state":"open","state_note":null,"filed_at":"2026-11-14T06:44:25.386Z","budget_usd":2.5,"ledger":{"action_id":"80203dd3-456f-4480-b518-3336d331f348","variant":"full","status":"open","cost_usd":2.5,"backing_usd":0},"work":[],"kept_by_this_node":true}],"requests_total":1,"subclaim_count":10,"instance_count":6,"source_map":{"summary":"Three of the six sources shown come from Anthropic, and two of those study the same trained model: the 2025 blind audit and the later automated rerun that isolates the SAE tools' contribution. The third, AuditBench, is a separate benchmark whose null result for its SAE tool cuts against the laboratory's own research investment. The model-diffing comparison comes from a different group that strengthened the non-SAE baseline, and the position paper reports no experiment of its own.","material":true,"sources_considered":5,"sources_read":4,"edges_recorded":1,"mapped_by":"claim_steward","mapped_at":"2026-11-14T06:45:16.660Z"},"subclaims":[{"id":"ef3572b3-9728-42bf-9cfb-1cc0bada2103","edge_id":"f90ec2aa-ea58-4e54-928f-05b8493b0409","text":"Sparse autoencoders are much less useful for uncovering a fine-tuned model's hidden objective when their training data omit the documents that describe the objective.","type":"empirical_derived","relation":"answers","reasoning":"A candidate condition separating audit wins from losses. Anthropic's automated-auditing study (2025) reran its hidden-objective game with SAEs trained on different data: an SAE trained on generic pretraining and chat data without the planted synthetic documents had no feature corresponding to the reward-model biases and the agent never won; an SAE trained on all three provided datasets also gave lower-quality features. AuditBench's SAE tool used a dictionary fit to the base model and gave no gain. Matcher found no existing node.","confidence":1,"assessment_status":null,"assessment_confidence":null,"assessment_credence":null,"coverage":{"examined":false},"argument_id":null,"argument_name":null},{"id":"391dac24-440f-438b-8cff-8933dfac302c","edge_id":"b87da4b1-9897-4852-bd19-b963ebb990a6","text":"In Anthropic's reward-model-sycophancy audit, sparse-autoencoder tools improve automated agents' objective-discovery success over chat and training-data search tools alone.","type":"empirical_derived","relation":"answers","reasoning":"The one located tool ablation that isolates the dictionary method's contribution inside the hidden-objective audit: the 2025 automated-auditing study ran the investigator agent with seven toolsets and reports interpretability tools lifting win rates from near 0% to 10 to 20% when all three datasets were provided, with a smaller gain when semantic search over the pretraining data alone was available. Matcher found no existing node.","confidence":1,"assessment_status":null,"assessment_confidence":null,"assessment_credence":null,"coverage":{"examined":false},"argument_id":null,"argument_name":null},{"id":"86063716-4ef1-48dc-abe6-02ef8b75f176","edge_id":"2bae5a08-7216-4331-b789-3baad200d28f","text":"In agent-based alignment auditing, black-box tools outperform white-box interpretability tools.","type":"empirical_verifiable","relation":"answers","reasoning":"The null-side answer for auditing: across 56 hidden-behaviour models, AuditBench found scaffolded black-box tools ahead of every activation-based tool. If it holds generally, dictionary methods do not outperform at this kind of discovery.","confidence":1,"assessment_status":"supported","assessment_confidence":0.6,"assessment_credence":0.6,"coverage":{"examined":true},"argument_id":null,"argument_name":null},{"id":"461a798c-19b2-4b7a-9f26-aa1dfed627fb","edge_id":"6e058872-a59d-4cdb-b2db-4f7c5f7919a7","text":"On AuditBench, logit-lens and sparse-autoencoder tools provide little benefit across tested target-model types.","type":"empirical_verifiable","relation":"answers","reasoning":"The most direct controlled measurement of an SAE tool's marginal contribution to finding unspecified behaviours: about 34.5% success with the SAE tool against 35.5% without, over 56 target models.","confidence":1,"assessment_status":"verified","assessment_confidence":0.85,"assessment_credence":0.88,"coverage":{"examined":true},"argument_id":null,"argument_name":null},{"id":"fc567967-00d4-456b-b196-d188db1187c4","edge_id":"d4ae54d2-e811-4e81-8f3a-8eab83e11c53","text":"Sparse autoencoders can reveal learned model knowledge when its application context differs semantically from training data.","type":"empirical_derived","relation":"answers","reasoning":"The condition proposed by the original hidden-objective audit for when SAE inspection beats semantic search of training data: when the context in which the model applies what it learned does not resemble the data that taught it.","confidence":1,"assessment_status":"supported","assessment_confidence":0.65,"assessment_credence":0.7,"coverage":{"examined":true},"argument_id":null,"argument_name":null},{"id":"4dbe502e-402c-4ae9-8c6f-57a252a0d324","edge_id":"79bb74ea-d68d-44d2-aee7-56a6daa6b7e5","text":"Sparse autoencoders are powerful tools for discovering previously unknown concepts represented in language models.","type":"evaluative","relation":"answers","reasoning":"The widely cited position that dictionary methods come into their own whenever the concept is not known in advance. It is the prominent affirmative answer and should be shown beside the controlled comparisons.","confidence":1,"assessment_status":"inconclusive","assessment_confidence":0.6,"assessment_credence":null,"coverage":{"examined":true},"argument_id":null,"argument_name":null},{"id":"846a4d41-3752-45ed-9210-b4177d75b5e9","edge_id":"160c4f8d-f3e7-4928-aef1-c89c4a7aee75","text":"In API-only model diffing, an improved LLM baseline performs comparably to SAE-based methods and surfaces more abstract behavioral differences.","type":"empirical_derived","relation":"answers","reasoning":"The null-side answer for model diffing from outputs: a language-model baseline that reads and compares outputs does about as well as the SAE-based method.","confidence":1,"assessment_status":"supported","assessment_confidence":0.75,"assessment_credence":0.7,"coverage":{"examined":true},"argument_id":null,"argument_name":null},{"id":"653e653c-cda0-4e4c-bfe2-de7f99acee34","edge_id":"ad6e2e1b-9318-4a2c-9bb1-8a7524332748","text":"SAE methods suit low-level model details, while LLM methods surface abstract behavioral differences.","type":"evaluative","relation":"answers","reasoning":"A candidate condition by type of property: dictionary methods for token-level and formatting differences, language-model summarisers for abstract behavioural ones.","confidence":1,"assessment_status":null,"assessment_confidence":null,"assessment_credence":null,"coverage":{"examined":false},"argument_id":null,"argument_name":null},{"id":"b9d866f3-2bda-456c-a10d-bc16f0450577","edge_id":"50b25408-8195-42aa-89b9-cc7ed72c4725","text":"Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE baselines and matching non-SAE methods.","type":"empirical_derived","relation":"answers","reasoning":"Bears on the 'related dictionary-learning methods' half for activation-level model diffing: a crosscoder variant is reported to beat SAE-based diffing and only match the non-dictionary activation-difference method on narrowly fine-tuned models.","confidence":1,"assessment_status":null,"assessment_confidence":null,"assessment_credence":null,"coverage":{"examined":false},"argument_id":null,"argument_name":null},{"id":"5393957c-bef5-4262-a152-84c5fe63ece2","edge_id":"d500ead0-8621-42fb-bb9b-12983786333d","text":"In Anthropic's 2025 blind auditing game, embedding-based search could have replicated the fastest team's SAE-based document discovery.","type":"empirical_verifiable","relation":"answers","reasoning":"The original audit's own concession about its headline SAE success: embedding search over the training data would have found the same documents. It marks the boundary of the claimed advantage.","confidence":1,"assessment_status":"supported","assessment_confidence":0.8,"assessment_credence":0.85,"coverage":{"examined":true},"argument_id":null,"argument_name":null}],"subclaims_assessed":6}