{"id":"4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073","title":"Authors propose ResCom to compensate for activation rounding in quantized language models","argument":"In a paper listed on arXiv, authors Ning Yang, Bohan Cui, Zhuowen Liu and Yuting Wang propose ResCom, a low-rank correction for large language models using low-bit weight–activation quantization. Quantization uses fewer bits to represent model values, reducing memory and computation, but rounding activations can still lower accuracy after calibration. ResCom uses the rounding residual—the difference introduced by rounding—to correct layer-output error. Its output-aware calibration chooses residual directions based on their contribution to reconstructing the output, then calculates the correction without gradient training or changing calibrated weights.\n\nOn Llama-3-8B with 4-bit weights and activations, the authors report that rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent. These are benchmark results, not evidence of improved performance in every deployment. The authors also report evaluations on models from 1B to 70B parameters and held-out C4. Their evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.\n\nHow well do the reported accuracy and memory results carry over to real-world deployments and workloads?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper's abstract, so it does not provide experimental details or enough information to assess the reported results independently. The reported benchmark results do not establish performance in real-world deployments or workloads.","speaker":{"name":"Yang, Ning, Cui, Bohan, Liu, Zhuowen, Wang, Yuting","roleAtSource":"Article author"},"source":{"title":"ResCom: Activation Residual Compensation for Large Language Model Quantization","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.35955","publishedAt":"2026-10-25","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors propose ResCom, a low-rank correction that uses activation rounding residuals to compensate for layer-output error in quantized language models.","excerpt":"We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"ResCom's output-aware calibration selects residual directions by their contribution to output reconstruction and computes the correction without gradient training or changes to calibrated weights.","excerpt":"selects residual directions by their contribution to output reconstruction and computes the correction in closed form, without gradient training or updates to the calibrated weights.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"On Llama-3-8B with 4-bit weights and activations, the authors report lower WikiText-2 perplexity, higher MMLU performance, and higher five-shot GSM8K accuracy with rank-32 attention corrections.","excerpt":"rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"The authors report evaluating models from 1B to 70B parameters and testing on held-out C4.","excerpt":"Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"The authors report that the evaluated FP32 implementation adds 2.25 percent to 4K prefill time, while a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","excerpt":"The evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.35955","checkedAt":"2026-11-06T15:32:08.736Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well do the reported accuracy and memory results carry over to real-world deployments and workloads?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Authors propose ResCom to compensate for activation rounding in quantized language models","summary":"In a paper listed on arXiv, authors Ning Yang, Bohan Cui, Zhuowen Liu and Yuting Wang propose ResCom, a low-rank correction for large language models using low-bit weight–activation quantization. Quantization uses fewer bits to represent model values, reducing memory and computation, but rounding activations can still lower accuracy after calibration. ResCom uses the rounding residual—the difference introduced by rounding—to correct layer-output error. Its output-aware calibration chooses residual directions based on their contribution to reconstructing the output, then calculates the correction without gradient training or changing calibrated weights.\n\nOn Llama-3-8B with 4-bit weights and activations, the authors report that rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent. These are benchmark results, not evidence of improved performance in every deployment. The authors also report evaluations on models from 1B to 70B parameters and held-out C4. Their evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.\n\nHow well do the reported accuracy and memory results carry over to real-world deployments and workloads?","question":"How well do the reported accuracy and memory results carry over to real-world deployments and workloads?","claims":[{"text":"The authors propose ResCom, a low-rank correction that uses activation rounding residuals to compensate for layer-output error in quantized language models.","excerpt":"We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.","support":"source-stated"},{"text":"ResCom's output-aware calibration selects residual directions by their contribution to output reconstruction and computes the correction without gradient training or changes to calibrated weights.","excerpt":"selects residual directions by their contribution to output reconstruction and computes the correction in closed form, without gradient training or updates to the calibrated weights.","support":"source-stated"},{"text":"On Llama-3-8B with 4-bit weights and activations, the authors report lower WikiText-2 perplexity, higher MMLU performance, and higher five-shot GSM8K accuracy with rank-32 attention corrections.","excerpt":"rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent.","support":"source-stated"},{"text":"The authors report evaluating models from 1B to 70B parameters and testing on held-out C4.","excerpt":"Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting.","support":"source-stated"},{"text":"The authors report that the evaluated FP32 implementation adds 2.25 percent to 4K prefill time, while a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","excerpt":"The evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","support":"source-stated"}],"limitations":["The supplied evidence is the paper's abstract, so it does not provide experimental details or enough information to assess the reported results independently.","The reported benchmark results do not establish performance in real-world deployments or workloads."],"mediaId":"c7221d8792126a940679437b5cfe9c93423e6907b6b61127350da8189d6143cd","correctionReason":null,"id":"4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073","publicationId":"a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b","version":1,"sourceUrl":"https://arxiv.org/abs/2610.35955","sourceTitle":"ResCom: Activation Residual Compensation for Large Language Model Quantization","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-25","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T04:32:34.813Z","createdAt":"2026-11-06T11:16:14.551Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"c7221d8792126a940679437b5cfe9c93423e6907b6b61127350da8189d6143cd","kind":"article","url":"https://arxiv.org/abs/2610.35955","sourceUrl":"https://arxiv.org/abs/2610.35955","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T00:02:28.102Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.35955","checkedAt":"2026-11-06T15:32:08.736Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"c7221d8792126a940679437b5cfe9c93423e6907b6b61127350da8189d6143cd","confidence":0.91,"provider":"typesafe","model":"jev-1.13.0","inputHash":"2bd6a0a49cdbd0f16978b5feeafa98e4a8e97188655424bcaf153ccc462e6fb4","decidedAt":"2026-10-30T00:02:30.067Z"},"provenance":{"textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observationId":"d9ad45b78dfd118b5fbdce655b167019b3903c7b33e49b909ccbb63a2ee350cc","retrievedAt":"2026-10-30T00:02:27.825Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"fe37993f-5329-4acf-8efa-269d754aae20"},"reviewer":{"model":"gpt-6-luna","runId":"9654b5e0-5a16-4b50-8f57-a7216e2607d3"}},"audit":{"status":"passed","reviewedAt":"2026-11-06T11:16:14.551Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Yang, Ning","url":null,"kind":"unknown"},{"name":"Cui, Bohan","url":null,"kind":"unknown"},{"name":"Liu, Zhuowen","url":null,"kind":"unknown"},{"name":"Wang, Yuting","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Yang, Ning","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Yang, Ning","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}},{"actor":{"name":"Cui, Bohan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Cui, Bohan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}},{"actor":{"name":"Liu, Zhuowen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Liu, Zhuowen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}},{"actor":{"name":"Wang, Yuting","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Yuting","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}}],"participationCoverage":{"total":4,"returned":4,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.35955","snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073/source"}},"canonicalUrl":"https://escapemeridian.com/readings/authors-propose-rescom-to-compensate-for-activation-rounding-in-4c7631f7","formats":{"html":"/readings/authors-propose-rescom-to-compensate-for-activation-rounding-in-4c7631f7","markdown":"/readings/authors-propose-rescom-to-compensate-for-activation-rounding-in-4c7631f7.md","json":"/readings/authors-propose-rescom-to-compensate-for-activation-rounding-in-4c7631f7.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-quantization","name":"Quantization","explanation":"Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.","referenceUrl":"https://developers.google.com/machine-learning/glossary#quantization"},{"id":"term-llm","name":"Large language model","explanation":"A language model with many learned numerical settings, used to process and generate language.","referenceUrl":"https://developers.google.com/machine-learning/glossary#large-language-model"},{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073","revision":"a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b","revisionKind":"publication-id","title":"Authors propose ResCom to compensate for activation rounding in quantized language models","url":"https://escapemeridian.com/readings/authors-propose-rescom-to-compensate-for-activation-rounding-in-4c7631f7"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.35955","revision":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","revisionKind":"sha256-source-text","title":"ResCom: Activation Residual Compensation for Large Language Model Quantization","url":"https://arxiv.org/abs/2610.35955"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073","title":"Authors propose ResCom to compensate for activation rounding in quantized language models","argument":"In a paper listed on arXiv, authors Ning Yang, Bohan Cui, Zhuowen Liu and Yuting Wang propose ResCom, a low-rank correction for large language models using low-bit weight–activation quantization. Quantization uses fewer bits to represent model values, reducing memory and computation, but rounding activations can still lower accuracy after calibration. ResCom uses the rounding residual—the difference introduced by rounding—to correct layer-output error. Its output-aware calibration chooses residual directions based on their contribution to reconstructing the output, then calculates the correction without gradient training or changing calibrated weights.\n\nOn Llama-3-8B with 4-bit weights and activations, the authors report that rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent. These are benchmark results, not evidence of improved performance in every deployment. The authors also report evaluations on models from 1B to 70B parameters and held-out C4. Their evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.\n\nHow well do the reported accuracy and memory results carry over to real-world deployments and workloads?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper's abstract, so it does not provide experimental details or enough information to assess the reported results independently. The reported benchmark results do not establish performance in real-world deployments or workloads.","speaker":{"name":"Yang, Ning, Cui, Bohan, Liu, Zhuowen, Wang, Yuting","roleAtSource":"Article author"},"source":{"title":"ResCom: Activation Residual Compensation for Large Language Model Quantization","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.35955","publishedAt":"2026-10-25","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors propose ResCom, a low-rank correction that uses activation rounding residuals to compensate for layer-output error in quantized language models.","excerpt":"We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"ResCom's output-aware calibration selects residual directions by their contribution to output reconstruction and computes the correction without gradient training or changes to calibrated weights.","excerpt":"selects residual directions by their contribution to output reconstruction and computes the correction in closed form, without gradient training or updates to the calibrated weights.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"On Llama-3-8B with 4-bit weights and activations, the authors report lower WikiText-2 perplexity, higher MMLU performance, and higher five-shot GSM8K accuracy with rank-32 attention corrections.","excerpt":"rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"The authors report evaluating models from 1B to 70B parameters and testing on held-out C4.","excerpt":"Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"},{"text":"The authors report that the evaluated FP32 implementation adds 2.25 percent to 4K prefill time, while a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","excerpt":"The evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","url":"https://arxiv.org/abs/2610.35955","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.35955","checkedAt":"2026-11-06T15:32:08.736Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well do the reported accuracy and memory results carry over to real-world deployments and workloads?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Authors propose ResCom to compensate for activation rounding in quantized language models","summary":"In a paper listed on arXiv, authors Ning Yang, Bohan Cui, Zhuowen Liu and Yuting Wang propose ResCom, a low-rank correction for large language models using low-bit weight–activation quantization. Quantization uses fewer bits to represent model values, reducing memory and computation, but rounding activations can still lower accuracy after calibration. ResCom uses the rounding residual—the difference introduced by rounding—to correct layer-output error. Its output-aware calibration chooses residual directions based on their contribution to reconstructing the output, then calculates the correction without gradient training or changing calibrated weights.\n\nOn Llama-3-8B with 4-bit weights and activations, the authors report that rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent. These are benchmark results, not evidence of improved performance in every deployment. The authors also report evaluations on models from 1B to 70B parameters and held-out C4. Their evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.\n\nHow well do the reported accuracy and memory results carry over to real-world deployments and workloads?","question":"How well do the reported accuracy and memory results carry over to real-world deployments and workloads?","claims":[{"text":"The authors propose ResCom, a low-rank correction that uses activation rounding residuals to compensate for layer-output error in quantized language models.","excerpt":"We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.","support":"source-stated"},{"text":"ResCom's output-aware calibration selects residual directions by their contribution to output reconstruction and computes the correction without gradient training or changes to calibrated weights.","excerpt":"selects residual directions by their contribution to output reconstruction and computes the correction in closed form, without gradient training or updates to the calibrated weights.","support":"source-stated"},{"text":"On Llama-3-8B with 4-bit weights and activations, the authors report lower WikiText-2 perplexity, higher MMLU performance, and higher five-shot GSM8K accuracy with rank-32 attention corrections.","excerpt":"rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent.","support":"source-stated"},{"text":"The authors report evaluating models from 1B to 70B parameters and testing on held-out C4.","excerpt":"Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting.","support":"source-stated"},{"text":"The authors report that the evaluated FP32 implementation adds 2.25 percent to 4K prefill time, while a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","excerpt":"The evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.","support":"source-stated"}],"limitations":["The supplied evidence is the paper's abstract, so it does not provide experimental details or enough information to assess the reported results independently.","The reported benchmark results do not establish performance in real-world deployments or workloads."],"mediaId":"c7221d8792126a940679437b5cfe9c93423e6907b6b61127350da8189d6143cd","correctionReason":null,"id":"4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073","publicationId":"a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b","version":1,"sourceUrl":"https://arxiv.org/abs/2610.35955","sourceTitle":"ResCom: Activation Residual Compensation for Large Language Model Quantization","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-25","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T04:32:34.813Z","createdAt":"2026-11-06T11:16:14.551Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"c7221d8792126a940679437b5cfe9c93423e6907b6b61127350da8189d6143cd","kind":"article","url":"https://arxiv.org/abs/2610.35955","sourceUrl":"https://arxiv.org/abs/2610.35955","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T00:02:28.102Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.35955","checkedAt":"2026-11-06T15:32:08.736Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"c7221d8792126a940679437b5cfe9c93423e6907b6b61127350da8189d6143cd","confidence":0.91,"provider":"typesafe","model":"jev-1.13.0","inputHash":"2bd6a0a49cdbd0f16978b5feeafa98e4a8e97188655424bcaf153ccc462e6fb4","decidedAt":"2026-10-30T00:02:30.067Z"},"provenance":{"textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observationId":"d9ad45b78dfd118b5fbdce655b167019b3903c7b33e49b909ccbb63a2ee350cc","retrievedAt":"2026-10-30T00:02:27.825Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"fe37993f-5329-4acf-8efa-269d754aae20"},"reviewer":{"model":"gpt-6-luna","runId":"9654b5e0-5a16-4b50-8f57-a7216e2607d3"}},"audit":{"status":"passed","reviewedAt":"2026-11-06T11:16:14.551Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Yang, Ning","url":null,"kind":"unknown"},{"name":"Cui, Bohan","url":null,"kind":"unknown"},{"name":"Liu, Zhuowen","url":null,"kind":"unknown"},{"name":"Wang, Yuting","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Yang, Ning","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Yang, Ning","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}},{"actor":{"name":"Cui, Bohan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Cui, Bohan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}},{"actor":{"name":"Liu, Zhuowen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Liu, Zhuowen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}},{"actor":{"name":"Wang, Yuting","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Yuting","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.35955","locator":"head/meta[name=author|citation_author]"},"assertionId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T15:32:08.736Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","sourceUrl":"https://arxiv.org/abs/2610.35955","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z"}}],"participationCoverage":{"total":4,"returned":4,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.35955","snapshotId":"bcdcfd56f7127821e91d85b7db39f833c55284e7f4dc0c4a3d9e5725d4eccd87","contentSha256":"392bdc19da328a174c2f430fc34bd902286672951475550bae04c553e942d2d4","textHash":"2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7","observedAt":"2026-11-06T15:32:08.736Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/4c7631f7bed976053a35cfa5d2200d13b61a24f48705818b7593fa66bdd6d073/source"}}}],"audits":[{"publicationId":"a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly attributes ResCom to the four listed authors in an arXiv paper. It keeps the benchmark and overhead figures framed as what the authors report, not as established fact.","excerpt":"Authors:Ning Yang, Bohan Cui, Zhuowen Liu, Yuting Wang"},{"dimension":"summary","status":"pass","explanation":"The title, summary and question match the abstract. Quantization and the rounding residual are explained in plain language, and the cost-versus-accuracy tradeoff is clear. Benchmark names (perplexity, MMLU, GSM8K) and terms such as low-rank, prefill and FP32/FP16 could get tooltips, but the reading stays understandable. A small optional repair would restore the abstract's motivating point that rounded inputs cannot distinguish activations in the same bin; its omission does not distort the meaning. The question is a genuine open issue.","excerpt":"Reconstruction from rounded inputs cannot distinguish activations within the same quantization bin."},{"dimension":"date","status":"pass","explanation":"The draft's originalPublished date of 2026-10-25 matches the v1 submission date on arXiv. The later feed date was correctly not used.","excerpt":"[Submitted on 25 Oct 2026]"},{"dimension":"support","status":"pass","explanation":"Every claim and number matches the abstract verbatim: 7.708 to 7.147 perplexity, plus 2.89 MMLU points, GSM8K 23.96 to 36.16, 2.25 percent prefill overhead and 27.7 percent memory reduction. The draft also keeps the distinction between the FP32 implementation and the separate compact-storage implementation. Its benchmark caveat is appropriate, and it does not claim results for the 1B–70B or C4 evaluations beyond what the abstract states.","excerpt":"The evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page, which is the direct original source. No other source media was found.","excerpt":null}],"createdAt":"2026-11-06T11:16:14.551Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}