Claims

Search the graph by meaning. Each result carries its current verdict; open one to see its decomposition, provenance, and the reasoning behind the assessment.

40 results

Methods and findings aimed at identifying and explaining the internal features, circuits, and computational mechanisms of neural networks, especially language models. Includes sparse autoencoders and dictionary-learning approaches; excludes interpretability of verbal chain-of-thought alone.

Delta-Crosscoder identifies and mitigates narrow fine-tuning behaviors, outperforming SAE baselines and matching non-SAE methods.

InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Large language modelsimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

When do sparse autoencoders and related dictionary-learning methods outperform other methods for discovering language-model properties?

LeaningThe evidence favors an answer without establishing it.constitution →QuestionA question: answered rather than affirmed or denied, with its candidate answers beneath it. Its standing says how well answered it is.constitution →Large language models Latent Knowledge in Language Modelsimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

In Anthropic's 2025 blind auditing game, embedding-based search could have replicated the fastest team's SAE-based document discovery.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Anthropic AI Capability Evaluationimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse probes on sufficiently good sparse autoencoder latents are less likely than dense probes to overfit spurious correlations.

DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Machine Learning Shortcut Learning Neural Network Generalizationimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Simple non-SAE baselines match SAE results on dataset-quality, spurious-correlation, and multi-token probing tasks.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Shortcut Learningimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Spurious correlates of a label are as likely as the target concept to be represented by individual sparse autoencoder latents.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Sparse Autoencoders Shortcut Learningimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Under what conditions, if any, do sparse-autoencoder and other dictionary-learning features outperform linear probes and difference-of-means steering on labeled language-model tasks?

LeaningThe evidence favors an answer without establishing it.constitution →QuestionA question: answered rather than affirmed or denied, with its candidate answers beneath it. Its standing says how well answered it is.constitution →Large language models Latent Knowledge in Language Modelsimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

With features selected and labelled by a supervised pipeline, SAE steering approaches LoRA performance on AxBench.

InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language models AI Capability Evaluationimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

In API-only model diffing, an improved LLM baseline performs comparably to SAE-based methods and surfaces more abstract behavioral differences.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Large language models Sparse Autoencodersimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

On AuditBench, logit-lens and sparse-autoencoder tools provide little benefit across tested target-model types.

VerifiedThe evidence, examined directly, establishes the claim.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AI Capability Evaluation Large language modelsimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse autoencoders are powerful tools for discovering previously unknown concepts represented in language models.

InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →evaluativeA judgment of worth or quality against some standard: good, fair, effective.constitution →Large language modelsimportance · notableImportance 0.55, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

In agent-based alignment auditing, black-box tools outperform white-box interpretability tools.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AI Alignment AI Safety AI Agentsimportance · notableImportance 0.55, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

With output-score feature selection, SAE steering is competitive with supervised steering methods on AxBench.

DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →AI Capability Evaluation Large language modelsimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Claude 3.5 Haiku represents a couplet's intended rhyme word before generating its second line.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language models Anthropic Latent Reasoning in Language Modelsimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Selecting SAE features by high output score improves their steering performance relative to unfiltered SAE features.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencoders LLM Steeringimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse autoencoders can reveal learned model knowledge when its application context differs semantically from training data.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Latent Knowledge in Language Models Large language modelsimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

On AxBench, SAE steering underperforms both prompting and difference-in-means representations.

VerifiedThe evidence, examined directly, establishes the claim.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AI Capability Evaluation AxBenchimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse autoencoders rarely outperform simple baselines on practical downstream interpretability tasks.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →importance · notableImportance 0.55, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse autoencoder probes do not consistently outperform baseline probes on language-model activation classification tasks.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Large language modelsimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Linear probes on model activations distinguish sandbagging from benign language models with low false-positive rates.

DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →AI Sandbagging Large language modelsimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

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.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencoders Mesa-Optimizationimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Arbitrary capabilities can be hidden in neural network weights so that, under cryptographic assumptions, no polynomial-time white-box auditor can elicit them.

SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →mathematicalA proposition of mathematics: true or false by proof rather than by observation. Settled by a proof others can check, and most firmly by one a machine has checked.constitution →Computational complexity theory Model Backdoor Attacksimportance · notableImportance 0.55, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

HypotheSAEs' benchmark gave NLParam and HypoGenic baselines more generous resource budgets than their published experiments.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →evaluativeA judgment of worth or quality against some standard: good, fair, effective.constitution →Sparse Autoencoders AI Capability Evaluationimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

On HEADLINES and CONGRESS, HypotheSAEs produce more hypotheses rated predictive, helpful, and interpretable than baselines.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Sparse Autoencoders AI Scientific Discovery SAEBenchimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

In the same hypothesis-generation pipeline, SAE features produce more significant hypotheses than embedding, PCA, or bottleneck features.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Sparse Autoencoders AI Scientific Discoveryimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse-autoencoder neurons in text embeddings receive higher-fidelity natural-language interpretations than raw, PCA, or bottleneck features.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Sparse Autoencoders Machine Learningimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Linear activation probes detect sandbagging more effectively than black-box methods across evaluation tasks.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →AI Sandbagging AI Capability Evaluation Latent Knowledge in Language Modelsimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

HypotheSAEs outperform BERTopic, NLParam, and HypoGenic at identifying reference hypotheses on synthetic datasets.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Sparse Autoencoders AI Capability Evaluationimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

API-only model diffing cannot surface behavioral differences that do not manifest in model outputs.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Machine Learning Latent Knowledge in Language Modelsimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

SAE-based and LLM-based API-only diffing both identify harmful advice and actions in risky-finance-finetuned Qwen-2.5-7B-Instruct.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language models Emergent Misalignmentimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Token-level model-diffing methods inherently struggle to generate abstract behavioral hypotheses.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Large language modelsimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

LLM-based model diffing yields higher hypothesis acceptance rates than SAE-based model diffing.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencoders Large language models Crosscodersimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Probing max-pooled SAE pre-activations performs about as well as probing max-pooled SAE latent activations.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencodersimportance · settledImportance 0.15, from 0 to 1 · settled: uncontested, so low even when much depends on it. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Task-prompted last-token activation probes are more data-efficient than SAE probes, while matching them with plentiful data.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencodersimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Token-wise max-pooling, rather than the SAE representation, explains degraded SAE probing on sentiment classification.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Sparse Autoencoders Sentiment Analysisimportance · minorImportance 0.25, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Max-pooling improves SAE probe performance but reduces raw-activation probe performance.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencodersimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Negative results on prespecified-concept tasks need not deter research into sparse autoencoders.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →normativeA claim about what should be done or how things ought to be, settled by argument rather than evidence alone.constitution →Sparse Autoencodersimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse autoencoders can discover concepts in text more effectively than keyword and topic models.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →evaluativeA judgment of worth or quality against some standard: good, fair, effective.constitution →Sparse Autoencoders Large language modelsimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Sparse autoencoders show positive results against baselines on known-concept tasks, especially on cost and efficiency.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →Sparse Autoencodersimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Unreliable sparse-autoencoder features are more tolerable for concept discovery than detection or steering because discoveries can be validated.

UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Sparse Autoencoders LLM Steeringimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

Contribute

If a claim here is wrong, or missing evidence, open it: every claim page carries its own entry for challenges, evidence, and corrections. If the graph is missing a claim or a question entirely, propose it below. A proposal is reviewed on its wording; accepted proposals are matched against the graph before they enter it.