{"paper":{"title":"Layer by Layer: Uncovering Hidden Representations in Language Models","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Intermediate layers in language models often encode richer representations than the final layer for downstream tasks.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Dan Zhao, Jalal Naghiyev, Md Rifat Arefin, Niket Patel, Oscar Skean, Ravid Shwartz-Ziv, Yann LeCun","submitted_at":"2025-02-04T05:03:42Z","abstract_excerpt":"From extracting features to generating text, the outputs of large language models (LLMs) typically rely on the final layers, following the conventional wisdom that earlier layers capture only low-level cues. However, our analysis shows that intermediate layers can encode even richer representations, often improving performance on a range of downstream tasks. To explain and quantify these hidden-layer properties, we propose a unified framework of representation quality metrics based on information theory, geometry, and invariance to input perturbations. Our framework highlights how each layer b"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"our analysis shows that intermediate layers can encode even richer representations, often improving performance on a range of downstream tasks... intermediate layers consistently provide stronger features, challenging the standard view on final-layer embeddings","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that the proposed metrics based on information theory, geometry, and invariance to input perturbations accurately capture representation quality relevant to downstream task performance","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Intermediate layers in LLMs consistently provide stronger features than final layers across tasks and architectures, as quantified by a new framework of information-theoretic, geometric, and invariance metrics.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Intermediate layers in language models often encode richer representations than the final layer for downstream tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"dc8b59d945f23b502183cf63228d93b678d15de36e79b0be330c750dd4d9f48f"},"source":{"id":"2502.02013","kind":"arxiv","version":2},"verdict":{"id":"e9508d18-ead8-4936-a86f-0f4c9811f422","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T16:25:43.473004Z","strongest_claim":"our analysis shows that intermediate layers can encode even richer representations, often improving performance on a range of downstream tasks... intermediate layers consistently provide stronger features, challenging the standard view on final-layer embeddings","one_line_summary":"Intermediate layers in LLMs consistently provide stronger features than final layers across tasks and architectures, as quantified by a new framework of information-theoretic, geometric, and invariance metrics.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that the proposed metrics based on information theory, geometry, and invariance to input perturbations accurately capture representation quality relevant to downstream task performance","pith_extraction_headline":"Intermediate layers in language models often encode richer representations than the final layer for downstream tasks."},"references":{"count":173,"sample":[{"doi":"","year":2022,"title":"Agrawal, K. K., Mondal, A. K., Ghosh, A., and Richards, B. - ReQ : Assessing representation quality in self-supervised learning by measuring eigenspectrum decay. NeurIPs, 2022","work_id":"8fe0fb88-86eb-414b-b26d-362bdcf0cf9e","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Alain, G. and Bengio, Y. Understanding intermediate layers using linear classifier probes. ICLR, 2017","work_id":"a916bb7d-1cf2-481f-88c9-5d8cc9240c10","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"R., Subbaraj, G., Gontier, N., LeCun, Y., Rish, I., Shwartz-Ziv, R., and Pal, C","work_id":"b4a8c8bf-2a6e-433d-8438-d1898900a6ac","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Information theory with kernel methods","work_id":"cd38bb25-b39e-4543-8a17-5c6e2db8d88b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"BeIT : Bert pre-training of image transformers","work_id":"c29e2625-5fb6-47c4-b15e-ad4380212f00","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":173,"snapshot_sha256":"4fc9c8e093cc20e5ad7db16f3600410d3e100fea8d14d137fc6435efbe234874","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"52234e52e1e85226baf11be515a74ec46aff0d131f1979ad36a5a9b8285dd307"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}