{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:EYQT6WP5QSAXINLYVQGPPNIL6Y","short_pith_number":"pith:EYQT6WP5","canonical_record":{"source":{"id":"2602.02001","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-02T12:02:21Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"a6fbbf441ccf1c51c5f7bb7fa1fe6b5154b06b56bf1542f5a55c4f1d06f9b0b5","abstract_canon_sha256":"15d52bcaf513c895067d8f1f1e4bd8f594452bc1631331d63ac116a542a2d39b"},"schema_version":"1.0"},"canonical_sha256":"26213f59fd8481743578ac0cf7b50bf60169859f1e1e0e069a2512ffb16fc557","source":{"kind":"arxiv","id":"2602.02001","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2602.02001","created_at":"2026-05-18T03:09:24Z"},{"alias_kind":"arxiv_version","alias_value":"2602.02001v2","created_at":"2026-05-18T03:09:24Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.02001","created_at":"2026-05-18T03:09:24Z"},{"alias_kind":"pith_short_12","alias_value":"EYQT6WP5QSAX","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"EYQT6WP5QSAXINLY","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"EYQT6WP5","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:EYQT6WP5QSAXINLYVQGPPNIL6Y","target":"record","payload":{"canonical_record":{"source":{"id":"2602.02001","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-02T12:02:21Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"a6fbbf441ccf1c51c5f7bb7fa1fe6b5154b06b56bf1542f5a55c4f1d06f9b0b5","abstract_canon_sha256":"15d52bcaf513c895067d8f1f1e4bd8f594452bc1631331d63ac116a542a2d39b"},"schema_version":"1.0"},"canonical_sha256":"26213f59fd8481743578ac0cf7b50bf60169859f1e1e0e069a2512ffb16fc557","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:24.016016Z","signature_b64":"Kv8x0UBvd8clA1J//OLMP5mrOa2kbIKrmxw4MPEu1WCjoq8RyaFx2rV3JMZ0jywtgQw7SNvlhiPr3mTHyzKqDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"26213f59fd8481743578ac0cf7b50bf60169859f1e1e0e069a2512ffb16fc557","last_reissued_at":"2026-05-18T03:09:24.015462Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:24.015462Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2602.02001","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:24Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"oyWY+MUE4UGYPWBuMM4j5crXU/3GChKnp2ce+J69k+K3RDz6ZQJcUaVKhnVAEOBNbGhASz3RXfBpejYxhKOdAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-01T04:37:02.907528Z"},"content_sha256":"b2c4950706d14f7c46e270d2f4aa2d050486198eccf8afdcf68f6c848fb508cc","schema_version":"1.0","event_id":"sha256:b2c4950706d14f7c46e270d2f4aa2d050486198eccf8afdcf68f6c848fb508cc"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:EYQT6WP5QSAXINLYVQGPPNIL6Y","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Preserve-Then-Quantize: Balancing Rank Budgets for Quantization Error Reconstruction in LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Allocating part of the rank budget to preserve top singular directions of activation-scaled weights before quantization allows better reconstruction of the remaining error in LLMs.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Albert No, Dongjae Jeon, Moongyu Jeon, Soeun Kim, Yoonjun Cho","submitted_at":"2026-02-02T12:02:21Z","abstract_excerpt":"Quantization Error Reconstruction (QER) reduces accuracy loss in Post-Training Quantization (PTQ) by approximating weights as $\\mathbf{W} \\approx \\mathbf{Q} + \\mathbf{L}\\mathbf{R}$, using a rank-$r$ correction to reconstruct quantization error. Prior methods devote the full rank budget to error reconstruction, which is suboptimal when $\\mathbf{W}$ has intrinsic low-rank structure and quantization corrupts dominant directions. We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-$k$ singular subspace of the activation-scaled weight before quant"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-k singular subspace of the activation-scaled weight before quantization, quantizes only the residual, and uses the remaining rank r-k for error reconstruction. Experiments demonstrate consistent perplexity reductions across diverse models and quantization settings in PTQ, along with a 5.9 percentage-point average gain on GLUE under 2-bit QPEFT.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That preserving the top-k singular subspace of the activation-scaled weight is optimal and that the theory-guided criterion for choosing k correctly balances quantization-exposed energy against unrecoverable error without model-specific overfitting.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SRR preserves the top-k singular subspace of activation-scaled weights before quantization, quantizes the residual, and allocates remaining rank for error reconstruction, yielding lower perplexity in PTQ and a 5.9-point GLUE gain in 2-bit QPEFT.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Allocating part of the rank budget to preserve top singular directions of activation-scaled weights before quantization allows better reconstruction of the remaining error in LLMs.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"808808eed0dd46f6c3652bcb8ec076f0417e3072ac9bca682e45374f76732b7f"},"source":{"id":"2602.02001","kind":"arxiv","version":2},"verdict":{"id":"fcc56da2-c2b5-423e-a5c8-9ed3f1e52bac","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T08:33:58.211026Z","strongest_claim":"We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-k singular subspace of the activation-scaled weight before quantization, quantizes only the residual, and uses the remaining rank r-k for error reconstruction. Experiments demonstrate consistent perplexity reductions across diverse models and quantization settings in PTQ, along with a 5.9 percentage-point average gain on GLUE under 2-bit QPEFT.","one_line_summary":"SRR preserves the top-k singular subspace of activation-scaled weights before quantization, quantizes the residual, and allocates remaining rank for error reconstruction, yielding lower perplexity in PTQ and a 5.9-point GLUE gain in 2-bit QPEFT.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That preserving the top-k singular subspace of the activation-scaled weight is optimal and that the theory-guided criterion for choosing k correctly balances quantization-exposed energy against unrecoverable error without model-specific overfitting.","pith_extraction_headline":"Allocating part of the rank budget to preserve top singular directions of activation-scaled weights before quantization allows better reconstruction of the remaining error in LLMs."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"436c8bb0698cd70b6e89e12991a31047311253bd22bad4e8b074555ecd31073a"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"fcc56da2-c2b5-423e-a5c8-9ed3f1e52bac"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T03:09:24Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"M6q/GDODN1+4BV1NQ0hz79QCpOoN+v6OMRgSLAkqB5SMBLrohjcKYfcG8q+/5riWhbJUQ1W8XH0fts0/lRMmAA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-01T04:37:02.907991Z"},"content_sha256":"3e1088472a98537472c2f06408177620a07736766f59db8e61473eb1b141cf3b","schema_version":"1.0","event_id":"sha256:3e1088472a98537472c2f06408177620a07736766f59db8e61473eb1b141cf3b"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y/bundle.json","state_url":"https://pith.science/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-01T04:37:02Z","links":{"resolver":"https://pith.science/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y","bundle":"https://pith.science/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y/bundle.json","state":"https://pith.science/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y/state.json","well_known_bundle":"https://pith.science/.well-known/pith/EYQT6WP5QSAXINLYVQGPPNIL6Y/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:EYQT6WP5QSAXINLYVQGPPNIL6Y","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"15d52bcaf513c895067d8f1f1e4bd8f594452bc1631331d63ac116a542a2d39b","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-02T12:02:21Z","title_canon_sha256":"a6fbbf441ccf1c51c5f7bb7fa1fe6b5154b06b56bf1542f5a55c4f1d06f9b0b5"},"schema_version":"1.0","source":{"id":"2602.02001","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2602.02001","created_at":"2026-05-18T03:09:24Z"},{"alias_kind":"arxiv_version","alias_value":"2602.02001v2","created_at":"2026-05-18T03:09:24Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.02001","created_at":"2026-05-18T03:09:24Z"},{"alias_kind":"pith_short_12","alias_value":"EYQT6WP5QSAX","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"EYQT6WP5QSAXINLY","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"EYQT6WP5","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:3e1088472a98537472c2f06408177620a07736766f59db8e61473eb1b141cf3b","target":"graph","created_at":"2026-05-18T03:09:24Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-k singular subspace of the activation-scaled weight before quantization, quantizes only the residual, and uses the remaining rank r-k for error reconstruction. Experiments demonstrate consistent perplexity reductions across diverse models and quantization settings in PTQ, along with a 5.9 percentage-point average gain on GLUE under 2-bit QPEFT."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That preserving the top-k singular subspace of the activation-scaled weight is optimal and that the theory-guided criterion for choosing k correctly balances quantization-exposed energy against unrecoverable error without model-specific overfitting."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"SRR preserves the top-k singular subspace of activation-scaled weights before quantization, quantizes the residual, and allocates remaining rank for error reconstruction, yielding lower perplexity in PTQ and a 5.9-point GLUE gain in 2-bit QPEFT."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Allocating part of the rank budget to preserve top singular directions of activation-scaled weights before quantization allows better reconstruction of the remaining error in LLMs."}],"snapshot_sha256":"808808eed0dd46f6c3652bcb8ec076f0417e3072ac9bca682e45374f76732b7f"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"436c8bb0698cd70b6e89e12991a31047311253bd22bad4e8b074555ecd31073a"},"paper":{"abstract_excerpt":"Quantization Error Reconstruction (QER) reduces accuracy loss in Post-Training Quantization (PTQ) by approximating weights as $\\mathbf{W} \\approx \\mathbf{Q} + \\mathbf{L}\\mathbf{R}$, using a rank-$r$ correction to reconstruct quantization error. Prior methods devote the full rank budget to error reconstruction, which is suboptimal when $\\mathbf{W}$ has intrinsic low-rank structure and quantization corrupts dominant directions. We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-$k$ singular subspace of the activation-scaled weight before quant","authors_text":"Albert No, Dongjae Jeon, Moongyu Jeon, Soeun Kim, Yoonjun Cho","cross_cats":["cs.AI"],"headline":"Allocating part of the rank budget to preserve top singular directions of activation-scaled weights before quantization allows better reconstruction of the remaining error in LLMs.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-02T12:02:21Z","title":"Preserve-Then-Quantize: Balancing Rank Budgets for Quantization Error Reconstruction in LLMs"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.02001","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T08:33:58.211026Z","id":"fcc56da2-c2b5-423e-a5c8-9ed3f1e52bac","model_set":{"reader":"grok-4.3"},"one_line_summary":"SRR preserves the top-k singular subspace of activation-scaled weights before quantization, quantizes the residual, and allocates remaining rank for error reconstruction, yielding lower perplexity in PTQ and a 5.9-point GLUE gain in 2-bit QPEFT.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Allocating part of the rank budget to preserve top singular directions of activation-scaled weights before quantization allows better reconstruction of the remaining error in LLMs.","strongest_claim":"We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-k singular subspace of the activation-scaled weight before quantization, quantizes only the residual, and uses the remaining rank r-k for error reconstruction. Experiments demonstrate consistent perplexity reductions across diverse models and quantization settings in PTQ, along with a 5.9 percentage-point average gain on GLUE under 2-bit QPEFT.","weakest_assumption":"That preserving the top-k singular subspace of the activation-scaled weight is optimal and that the theory-guided criterion for choosing k correctly balances quantization-exposed energy against unrecoverable error without model-specific overfitting."}},"verdict_id":"fcc56da2-c2b5-423e-a5c8-9ed3f1e52bac"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:b2c4950706d14f7c46e270d2f4aa2d050486198eccf8afdcf68f6c848fb508cc","target":"record","created_at":"2026-05-18T03:09:24Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"15d52bcaf513c895067d8f1f1e4bd8f594452bc1631331d63ac116a542a2d39b","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-02-02T12:02:21Z","title_canon_sha256":"a6fbbf441ccf1c51c5f7bb7fa1fe6b5154b06b56bf1542f5a55c4f1d06f9b0b5"},"schema_version":"1.0","source":{"id":"2602.02001","kind":"arxiv","version":2}},"canonical_sha256":"26213f59fd8481743578ac0cf7b50bf60169859f1e1e0e069a2512ffb16fc557","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"26213f59fd8481743578ac0cf7b50bf60169859f1e1e0e069a2512ffb16fc557","first_computed_at":"2026-05-18T03:09:24.015462Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T03:09:24.015462Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"Kv8x0UBvd8clA1J//OLMP5mrOa2kbIKrmxw4MPEu1WCjoq8RyaFx2rV3JMZ0jywtgQw7SNvlhiPr3mTHyzKqDA==","signature_status":"signed_v1","signed_at":"2026-05-18T03:09:24.016016Z","signed_message":"canonical_sha256_bytes"},"source_id":"2602.02001","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:b2c4950706d14f7c46e270d2f4aa2d050486198eccf8afdcf68f6c848fb508cc","sha256:3e1088472a98537472c2f06408177620a07736766f59db8e61473eb1b141cf3b"],"state_sha256":"cebcc58a8d4c3faea71a0a8d4c6563a388c1a4b256c367e31e227dcaa362b17e"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"owZHhommdfY/YVPXfIX7NYNTrMF/hGqwXHbsXS/B6uPFwqhN3yar77Z5BWXgbrqn+0s90JVfOFYAWXqqqAP2Bw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-01T04:37:02.911092Z","bundle_sha256":"bc366f4f70b5e23d8f88164a286db3fa2d687fa0f65a1b1d1a650a00363104c4"}}