{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "burns-et-al-contrast-consistent-search-2022", "title": "Burns and colleagues: contrast-consistent search for latent model knowledge", "summary": "Learned probes from internal model representations using consistency constraints between contrasting statements rather than truth labels, investigating whether latent answer information can be recovered when prompted outputs are unreliable.", "boundary": "The representation may contain useful predictive information without encoding a universal truth variable. Reported evaluation resolves the binary label orientation using labels; an unlabeled orientation proposal is not the same as a demonstrated label-free truth oracle.", "area": "Language-model representation analysis", "topic": "Artificial intelligence", "year": 2022, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "burns-et-al-contrast-consistent-search-2022", "target": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "humans": [ "Collin Burns", "Haotian Ye", "Dan Klein", "Jacob Steinhardt" ], "authors_complete": true, "human_role_note": "Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-2212-03827v1", "title": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1", "date": "2022-12-07" } ], "points": "55000", "published_grade": { "id": "burns-et-al-contrast-consistent-search-2022", "entry_id": "burns-et-al-contrast-consistent-search-2022", "band": "B", "modifier": "+", "target": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "what_changed": "Learned probes from internal model representations using consistency constraints between contrasting statements rather than truth labels, investigating whether latent answer information can be recovered when prompted outputs are unreliable.", "scope": "Learned probes from internal model representations using consistency constraints between contrasting statements rather than truth labels, investigating whether latent answer information can be recovered when prompted outputs are unreliable. The representation may contain useful predictive information without encoding a universal truth variable. Reported evaluation resolves the binary label orientation using labels; an unlabeled orientation proposal is not the same as a demonstrated label-free truth oracle.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Collin Burns", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Haotian Ye", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Dan Klein", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jacob Steinhardt", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The representation may contain useful predictive information without encoding a universal truth variable. Reported evaluation resolves the binary label orientation using labels; an unlabeled orientation proposal is not the same as a demonstrated label-free truth oracle.", "domain": "Language-model representation analysis" }, "sources": [ { "citation": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1" } ], "statement": "# Burns and colleagues: contrast-consistent search for latent model knowledge\n\n## A consistent answer can still be wrong\n\nIf a model says that a statement and its negation are both true, something has gone wrong. Contrast-consistent search uses relationships between contrasting statements to fit a small readout from a model's internal representations. A readout, often called a probe, maps internal numbers to a prediction. The question is whether useful answer information can be recovered even when a model's prompted response is unreliable.\n\n## Why the label direction is a real problem\n\nHere is an invented classroom puzzle. Two boxes always receive opposite labels, red and blue. You discover a rule that assigns the labels consistently, but nobody tells you which color means true. Swapping every red and blue label preserves the same opposite-label relationship. Consistency alone has not resolved the meaning of the labels.\n\n## Training and evaluation must be separated\n\nThe paper fits the probe without truth labels. Its reported evaluation, however, compares both binary label orientations and uses the better test accuracy. Those are different facts. An unlabeled fitting objective can be useful without establishing that a deployed system knows which orientation corresponds to truth. The paper discusses a further proposal for identifying the direction; that proposal should not be confused with the reported evaluation protocol.\n\n## What this result adds\n\nThe contribution is a concrete probe-fitting approach with empirical tests of information recovery and transfer. It does not prove that every model has a single truth variable, that logical consistency guarantees factual accuracy, or that the method detects an actively deceptive system. The distinction between available predictive information and a guaranteed truth detector is central to reading the work fairly.\n\n## A small check anyone can do\n\nWrite four harmless statements together with their negations. Make a toy table with consistent but deliberately swapped labels. Verify that the table passes an opposite-label consistency check while failing your intended truth labels. This is a demonstration of an ambiguity, not a reproduction of the research. Then inspect where the original experiment obtains the information needed to score its predictions.\n\n## Scientific scope\n\nThe representation may contain useful predictive information without encoding a universal truth variable. Reported evaluation resolves the binary label orientation using labels; an unlabeled orientation proposal is not the same as a demonstrated label-free truth oracle.\n\n## Sources for this explanation\n\n- [Discovering Latent Knowledge in Language Models Without Supervision](https://arxiv.org/abs/2212.03827v1): Pinned original contribution and stated hypotheses or experimental limits.\n", "summary": "Learned probes from internal model representations using consistency constraints between contrasting statements rather than truth labels, investigating whether latent answer information can be recovered when prompted outputs are unreliable.", "title": "Burns and colleagues: contrast-consistent search for latent model knowledge" }, "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "source_ref": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:1afa3f10cacf08d93fd5b3e516d8bfdb88a47f6248a58c16b056276a84051545" }, "target": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "burns-et-al-contrast-consistent-search-2022", "entry_id": "burns-et-al-contrast-consistent-search-2022", "revision": 1, "title": "Burns and colleagues: contrast-consistent search for latent model knowledge", "context": "Language-model representation analysis", "track": "Human-designed model analysis or safeguard method", "band": "B", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Learned probes from internal model representations using consistency constraints between contrasting statements rather than truth labels, investigating whether latent answer information can be recovered when prompted outputs are unreliable.", "scope": "Learned probes from internal model representations using consistency constraints between contrasting statements rather than truth labels, investigating whether latent answer information can be recovered when prompted outputs are unreliable. The representation may contain useful predictive information without encoding a universal truth variable. Reported evaluation resolves the binary label orientation using labels; an unlabeled orientation proposal is not the same as a demonstrated label-free truth oracle.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The representation may contain useful predictive information without encoding a universal truth variable. Reported evaluation resolves the binary label orientation using labels; an unlabeled orientation proposal is not the same as a demonstrated label-free truth oracle.", "references": [ { "title": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred." } ], "target": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "55000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "burns-et-al-contrast-consistent-search-2022", "target": { "record_id": "urn:uuid:ffd1041b-78ed-41b3-acde-e38ca04d6b3c", "version": 1, "content_hash": "sha256:9c81119cb906cc54a0e7f0f9622fc852b818bfa9ac455b9d0b9d191d41637636" }, "sections": [ { "title": "A consistent answer can still be wrong", "text": "If a model says that a statement and its negation are both true, something has gone wrong. Contrast-consistent search uses relationships between contrasting statements to fit a small readout from a model's internal representations. A readout, often called a probe, maps internal numbers to a prediction. The question is whether useful answer information can be recovered even when a model's prompted response is unreliable." }, { "title": "Why the label direction is a real problem", "text": "Here is an invented classroom puzzle. Two boxes always receive opposite labels, red and blue. You discover a rule that assigns the labels consistently, but nobody tells you which color means true. Swapping every red and blue label preserves the same opposite-label relationship. Consistency alone has not resolved the meaning of the labels." }, { "title": "Training and evaluation must be separated", "text": "The paper fits the probe without truth labels. Its reported evaluation, however, compares both binary label orientations and uses the better test accuracy. Those are different facts. An unlabeled fitting objective can be useful without establishing that a deployed system knows which orientation corresponds to truth. The paper discusses a further proposal for identifying the direction; that proposal should not be confused with the reported evaluation protocol." }, { "title": "What this result adds", "text": "The contribution is a concrete probe-fitting approach with empirical tests of information recovery and transfer. It does not prove that every model has a single truth variable, that logical consistency guarantees factual accuracy, or that the method detects an actively deceptive system. The distinction between available predictive information and a guaranteed truth detector is central to reading the work fairly." }, { "title": "A small check anyone can do", "text": "Write four harmless statements together with their negations. Make a toy table with consistent but deliberately swapped labels. Verify that the table passes an opposite-label consistency check while failing your intended truth labels. This is a demonstration of an ambiguity, not a reproduction of the research. Then inspect where the original experiment obtains the information needed to score its predictions." } ], "references": [ { "title": "Discovering Latent Knowledge in Language Models Without Supervision", "url": "https://arxiv.org/abs/2212.03827v1", "locator": "Pinned original contribution and stated hypotheses or experimental limits." } ] }, "reading_status": "EXPLANATORY_GUIDE" }