{"paper":{"title":"Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Depth registers with a hinge loss control residual reader magnitudes in SwiGLU, unlocking W4A4 quantization performance while leaving the w2 bilinear generator as the dominant error.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Ziyang Liu","submitted_at":"2026-04-20T11:47:11Z","abstract_excerpt":"We study post-training W4A4 quantization in a controlled 300M-parameter SwiGLU decoder-only language model trained on 5B tokens of FineWeb-Edu, and ask which input-activation sites dominate the error. Naive round-to-nearest W4A4 collapses validation perplexity from FP16 23.6 to 1727. A simple residual-axis training-time intervention -- Depth Registers with a register-magnitude hinge loss (DR+sink) -- reduces this to 119 (about 14x) at matched FP16 PPL and matched zero-shot capacity, and composes with SmoothQuant to 39.9 PPL. The residual ~2 PPL gap to FP16 is the diagnostic core. We decompose "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Elementary norm arguments show residual-axis magnitude control bounds readers tightly but leaves w2's bilinear input bounded only by the trivial product of factor bounds; empirically, DR+sink collapses reader kurtosis while leaving generators essentially unchanged, and the reader-rescued W4A4 residue is flat at ~0.28 nats across three matched checkpoints with Delta-remove(w2) dominating.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The partition into residual-axis readers (qkv, w1, w3) and block-internal generators (o_proj, w2) plus the claim that the hinge loss isolates the reader effect without confounding the generator behavior; the paper states its experiments do not isolate the partition from the hinge.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Depth Registers plus hinge loss cut W4A4-induced perplexity collapse from 1727 to 119 in a 300M SwiGLU model by selectively taming reader-layer activations while leaving bilinear generator tails largely untouched.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Depth registers with a hinge loss control residual reader magnitudes in SwiGLU, unlocking W4A4 quantization performance while leaving the w2 bilinear generator as the dominant error.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"58ab85e1b0d181e4b5c349833488fac87f801887f8ab34e4a25c0b609b88b776"},"source":{"id":"2604.18128","kind":"arxiv","version":2},"verdict":{"id":"dc7bc49f-6ff8-4702-9fde-af6c2a5b0061","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T04:57:30.913578Z","strongest_claim":"Elementary norm arguments show residual-axis magnitude control bounds readers tightly but leaves w2's bilinear input bounded only by the trivial product of factor bounds; empirically, DR+sink collapses reader kurtosis while leaving generators essentially unchanged, and the reader-rescued W4A4 residue is flat at ~0.28 nats across three matched checkpoints with Delta-remove(w2) dominating.","one_line_summary":"Depth Registers plus hinge loss cut W4A4-induced perplexity collapse from 1727 to 119 in a 300M SwiGLU model by selectively taming reader-layer activations while leaving bilinear generator tails largely untouched.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The partition into residual-axis readers (qkv, w1, w3) and block-internal generators (o_proj, w2) plus the claim that the hinge loss isolates the reader effect without confounding the generator behavior; the paper states its experiments do not isolate the partition from the hinge.","pith_extraction_headline":"Depth registers with a hinge loss control residual reader magnitudes in SwiGLU, unlocking W4A4 quantization performance while leaving the w2 bilinear generator as the dominant error."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.18128/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-20T04:24:53.568733Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"60d0cd2a3d90c58f67dbcb97c7815575e54dfaddd1b0ca01c2056042a8f477dd"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}