Ozdayi, Mustafa Safa, Wadhwa, Somin · Oct 12, 2026
Ozdayi and Wadhwa propose Artifact Distillation for transferring problem-solving procedures
Markdown · JSON · Original source
Mustafa Safa Ozdayi and Somin Wadhwa propose Artifact Distillation as an alternative to a smaller language model on a ’s outputs. Instead, the frontier model solves training instances and records its problem-solving procedure in an artifact bank, which the smaller model can query at inference without gradient updates. In the authors’ reported tests across four LiveBench tasks, an artifact bank written by Opus 4.5 let Claude Haiku 4.5 outperform unaided Claude Sonnet 4.5 on spatial reasoning, at 60% of its cost. On table joining, Haiku reached Sonnet’s level at 89% of its cost. The same raw bank also raised Qwen3.5-27B’s pooled score, suggesting that transfer may extend beyond the teacher’s model family. The approach shifts transfer from training the student to giving it reusable problem-solving procedures to consult. Whether the reported task results hold beyond these tests, and how the cost comparison changes in other settings, remains unknown. How well does Artifact Distillation transfer across tasks and settings beyond the four LiveBench tasks reported here?
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 12, 2026 (2026-10-12).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T05:22:58.287Z. First observed: 2026-10-30T06:42:51.154Z. Retrieved: 2026-10-31T08:33:37.446Z.
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 evidence provided is the paper’s abstract; it does not include the methods or results tables needed to assess the experiments in detail. The reported results cover four LiveBench tasks. The abstract does not establish how well the approach performs on other tasks or in other settings. The abstract reports cost comparisons but does not specify the cost accounting details. The comparisons should not be treated as evidence of equivalent savings in other deployments.
Open question
How well does Artifact Distillation transfer across tasks and settings beyond the four LiveBench tasks reported here?
Source-stated
Ozdayi and Wadhwa propose having a frontier model solve training instances and record its procedure in an artifact bank, which a smaller model queries at inference time without gradient updates.
a student queries at inference time, with no gradient updates
Source-stated
Across four LiveBench tasks, the authors report that an artifact bank written by Claude Opus 4.5 let Claude Haiku 4.5 outperform unaided Claude Sonnet 4.5 on spatial reasoning at 60% of its cost, and reach Sonnet’s level on table joining at 89% of its cost.
outperform unaided Claude Sonnet 4.5 on spatial reasoning (0.707 vs. 0.647) at 60% of its cost, and reach Sonnet 4.5's level on table joining at 89% of its cost
Source-stated
The authors report that the same raw artifact bank raised Qwen3.5-27B’s pooled score, indicating transfer beyond the teacher’s model family in their test.
The same raw bank raises Qwen3.5-27B's pooled score from 0.730 to 0.800, showing that transfer is not confined to the teacher's model family.
Source and contributor context
Ozdayi, Mustafa Safa, Wadhwa, Somin · 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.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
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.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 42c5586c-56bc-4400-8765-35e3b9ef21d8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T05:22:58.287Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6786afa6b1d5f1f2cb9d4562b199b9adab9d8a3f6d65cb6e831deb0369394a3d. Source text hash: f9c014f823fe56e3ce928b174c748d35c36f45cbf4950265c3694ee6897351a2.
Revision and source-check history
Version 1 · 2026-11-11T05:22:58.287Z
Ozdayi and Wadhwa propose Artifact Distillation for transferring problem-solving procedures
Mustafa Safa Ozdayi and Somin Wadhwa propose Artifact Distillation as an alternative to a smaller language model on a ’s outputs. Instead, the frontier model solves training instances and records its problem-solving procedure in an artifact bank, which the smaller model can query at inference without gradient updates. In the authors’ reported tests across four LiveBench tasks, an artifact bank written by Opus 4.5 let Claude Haiku 4.5 outperform unaided Claude Sonnet 4.5 on spatial reasoning, at 60% of its cost. On table joining, Haiku reached Sonnet’s level at 89% of its cost. The same raw bank also raised Qwen3.5-27B’s pooled score, suggesting that transfer may extend beyond the teacher’s model family. The approach shifts transfer from training the student to giving it reusable problem-solving procedures to consult. Whether the reported task results hold beyond these tests, and how the cost comparison changes in other settings, remains unknown. How well does Artifact Distillation transfer across tasks and settings beyond the four LiveBench tasks reported here?
Original source
cs.AI · Article
Source details
Original source date: Oct 12, 2026 (2026-10-12).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T05:22:58.287Z. First observed: 2026-10-30T06:42:51.154Z. Retrieved: 2026-10-31T08:33:37.446Z.
Limits
The evidence provided is the paper’s abstract; it does not include the methods or results tables needed to assess the experiments in detail. The reported results cover four LiveBench tasks. The abstract does not establish how well the approach performs on other tasks or in other settings. The abstract reports cost comparisons but does not specify the cost accounting details. The comparisons should not be treated as evidence of equivalent savings in other deployments.
Open question
How well does Artifact Distillation transfer across tasks and settings beyond the four LiveBench tasks reported here?
Source-stated
Ozdayi and Wadhwa propose having a frontier model solve training instances and record its procedure in an artifact bank, which a smaller model queries at inference time without gradient updates.
a student queries at inference time, with no gradient updates
Source-stated
Across four LiveBench tasks, the authors report that an artifact bank written by Claude Opus 4.5 let Claude Haiku 4.5 outperform unaided Claude Sonnet 4.5 on spatial reasoning at 60% of its cost, and reach Sonnet’s level on table joining at 89% of its cost.
outperform unaided Claude Sonnet 4.5 on spatial reasoning (0.707 vs. 0.647) at 60% of its cost, and reach Sonnet 4.5's level on table joining at 89% of its cost
Source-stated
The authors report that the same raw artifact bank raised Qwen3.5-27B’s pooled score, indicating transfer beyond the teacher’s model family in their test.
The same raw bank raises Qwen3.5-27B's pooled score from 0.730 to 0.800, showing that transfer is not confined to the teacher's model family.
Source and contributor context
Ozdayi, Mustafa Safa, Wadhwa, Somin · 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.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
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.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 42c5586c-56bc-4400-8765-35e3b9ef21d8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T05:22:58.287Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6786afa6b1d5f1f2cb9d4562b199b9adab9d8a3f6d65cb6e831deb0369394a3d. Source text hash: f9c014f823fe56e3ce928b174c748d35c36f45cbf4950265c3694ee6897351a2.
uncertain · 2026-11-11T05:22:58.287Z · anthropic/claude-opus-5.5
Publication: 42c5586c-56bc-4400-8765-35e3b9ef21d8
attribution — pass: The reading correctly credits the proposal to Mustafa Safa Ozdayi and Somin Wadhwa, the arXiv authors. It frames the results as the authors' reported tests rather than independently verified facts.
Authors:Mustafa Safa Ozdayi, Somin Wadhwa
summary — uncertain: The method and results are rendered faithfully, but the reading drops the authors' stated reason for the method. They argue that conventional distillation binds transfer to training compute and leaves transferred knowledge largely opaque. As a result, the 'why it matters' line is thin. Smallest repair: add one attributed sentence giving this motivation. Also, 'inference time' is never explained in plain language (for example, 'when the small model answers a question'), and the closing question is repeated inside the summary.
which binds transfer to training compute and leaves the transferred knowledge largely opaque
date — pass: originalPublished is 2026-10-12, which matches the v1 submission date on arXiv. The later feed date (2026-10-31) is correctly not used.
[Submitted on 12 Oct 2026]
support — pass: All claims match the abstract, including the 60% and 89% cost figures and the Qwen3.5-27B gain. Softening the authors' 'showing' to 'suggesting ... may' is cautious, but it is acceptable and keeps the source's assertion attributed. The limitations and the closing question follow from the four-task scope; neither invents benefits.
The same raw bank raises Qwen3.5-27B's pooled score from 0.730 to 0.800, showing that transfer is not confined to the teacher's model family.
media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Further training of an existing model to adapt it to a particular task or use.
Definition referenceOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivFurther training of an existing model to adapt it to a particular task or use.
Definition referenceOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv