Yang, Ning, Cui, Bohan, Liu, Zhuowen, Wang, Yuting · Oct 25, 2026
Authors propose ResCom to compensate for activation rounding in quantized language models
Markdown · JSON · Original source
In a paper listed on , authors Ning Yang, Bohan Cui, Zhuowen Liu and Yuting Wang propose ResCom, a low-rank correction for using low-bit weight–activation . 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. On 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 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 ; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16. How well do the reported accuracy and memory results carry over to real-world deployments and workloads?
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.
Original source
cs.AI · Article
Source details
Original source date: Oct 25, 2026 (2026-10-25).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-06T11:16:14.551Z. First observed: 2026-10-28T04:32:34.813Z. Retrieved: 2026-10-30T00:02:27.825Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
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 results do not establish performance in real-world deployments or workloads.
Open question
How well do the reported accuracy and memory results carry over to real-world deployments and workloads?
Source-stated
The authors propose ResCom, a low-rank correction that uses activation rounding residuals to compensate for layer-output error in quantized language models.
We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.
Source-stated
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.
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.
Source-stated
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.
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.
Source-stated
The authors report evaluating models from 1B to 70B parameters and testing on held-out C4.
Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting.
Source-stated
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.
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.
Source and contributor context
Yang, Ning, Cui, Bohan, Liu, Zhuowen, Wang, Yuting · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-06T11:16:14.551Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: d9ad45b78dfd118b5fbdce655b167019b3903c7b33e49b909ccbb63a2ee350cc. Source text hash: 2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7.
Revision and source-check history
Version 1 · 2026-11-06T11:16:14.551Z
Authors propose ResCom to compensate for activation rounding in quantized language models
In a paper listed on , authors Ning Yang, Bohan Cui, Zhuowen Liu and Yuting Wang propose ResCom, a low-rank correction for using low-bit weight–activation . 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. On 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 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 ; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16. How well do the reported accuracy and memory results carry over to real-world deployments and workloads?
Original source
cs.AI · Article
Source details
Original source date: Oct 25, 2026 (2026-10-25).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-06T11:16:14.551Z. First observed: 2026-10-28T04:32:34.813Z. Retrieved: 2026-10-30T00:02:27.825Z.
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 results do not establish performance in real-world deployments or workloads.
Open question
How well do the reported accuracy and memory results carry over to real-world deployments and workloads?
Source-stated
The authors propose ResCom, a low-rank correction that uses activation rounding residuals to compensate for layer-output error in quantized language models.
We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.
Source-stated
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.
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.
Source-stated
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.
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.
Source-stated
The authors report evaluating models from 1B to 70B parameters and testing on held-out C4.
Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting.
Source-stated
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.
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.
Source and contributor context
Yang, Ning, Cui, Bohan, Liu, Zhuowen, Wang, Yuting · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-06T11:16:14.551Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: d9ad45b78dfd118b5fbdce655b167019b3903c7b33e49b909ccbb63a2ee350cc. Source text hash: 2db9d2aa93e1c7bd1d3f3f6c07561dd06bc576f2ab3946a2c13f1acff3cddff7.
passed · 2026-11-06T11:16:14.551Z · anthropic/claude-opus-5.5
Publication: a0e0f0fc-3b2e-4e97-8e19-d16795dfcb3b
attribution — pass: 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.
Authors:Ning Yang, Bohan Cui, Zhuowen Liu, Yuting Wang
summary — pass: 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.
Reconstruction from rounded inputs cannot distinguish activations within the same quantization bin.
date — pass: The draft's originalPublished date of 2026-10-25 matches the v1 submission date on arXiv. The later feed date was correctly not used.
[Submitted on 25 Oct 2026]
support — pass: 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.
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.
media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. No other source media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference