Wu, Junxi, Guo, Yiwen, Xia, Shu-Tao · Oct 28, 2026
Wu, Guo and Xia propose input-adaptive computation for frozen LLMs
Markdown · JSON · Original source
Junxi Wu, Yiwen Guo and Shu-Tao Xia introduce SelfMorph, a method that lets a frozen choose how to organize its internal computation for each prompt. Rather than only generating more tokens, it selects among the original forward path, serial reuse of a layer span, or local branching and merging of perturbed hidden states before generating a response. The authors say these options have complementary benefits, so a lightweight selector reads the model’s depth-wise hidden state trajectory on the prompt and chooses a computation shape. They report an 8.06% mean relative improvement on five in-distribution and 4.45% on four unseen benchmarks, with the strongest performance in 24 of 27 settings. The idea is to adapt not just how much computation a model uses, but its structure. The abstract reports benchmark results, not evidence about performance or cost in deployed systems. How well does the selector’s benchmark advantage hold up across real-world tasks and ?
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 28, 2026 (2026-10-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T03:35:40.411Z. First observed: 2026-10-30T06:45:35.314Z. Retrieved: 2026-10-31T05:17:56.337Z.
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 abstract, which does not describe tasks, comparison methods, or how computational cost was measured. The abstract reports benchmark evaluations, not results from deployed systems or real-world tasks. Term-library review may be needed for SelfMorph, test- scaling, depth-wise hidden state trajectory, and other technical terms in the reading.
Open question
How well does the selector’s advantage hold up across real-world tasks and ?
Source-stated
The authors introduce SelfMorph, which makes the shape of a frozen LLM’s internal computation an input-dependent inference-time choice.
We introduce SelfMorph, which makes computational shape an input-dependent inference-time decision.
Source-stated
SelfMorph selects among the original forward path, serial reuse of a layer span, and local branching and merging of perturbed hidden states.
It unifies the original forward path (BASE), serial reuse of a layer span (REPEAT), and local branching and merging of perturbed hidden states (BRANCH)
Source-stated
The authors report an 8.06% mean relative improvement on five in-distribution benchmarks and 4.45% on four unseen benchmarks.
SelfMorph achieves a 8.06% mean relative improvement on five in-distribution benchmarks and 4.45% on four unseen benchmarks.
Source-stated
The authors report that SelfMorph had the strongest performance in 24 of 27 settings.
It achieves the strongest performance in 24 of 27 settings.
Escape Meridian interpretation
The abstract reports benchmark results, but does not establish performance or cost in deployed systems.
Evaluated on three LLMs spanning 4B-27B, SelfMorph achieves a 8.06% mean relative improvement on five in-distribution benchmarks and 4.45% on four unseen benchmarks.
Source and contributor context
Wu, Junxi, Guo, Yiwen, Xia, Shu-Tao · 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.
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.
Model inference
Using a trained model to produce a prediction or response to new input.
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: 41ddf19b-ea57-4642-8d7b-6ecf0eb8cc7a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T03:35:40.411Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 128846a9a1ee387cbc904afa61d0734778eb48ac3e2238a0989c3b652943c5f1. Source text hash: d9d839713d738db8550cd101fc98e0b741560f915f88c2bdc1d682eb430ba007.
Revision and source-check history
Version 1 · 2026-11-13T03:35:40.411Z
Wu, Guo and Xia propose input-adaptive computation for frozen LLMs
Junxi Wu, Yiwen Guo and Shu-Tao Xia introduce SelfMorph, a method that lets a frozen choose how to organize its internal computation for each prompt. Rather than only generating more tokens, it selects among the original forward path, serial reuse of a layer span, or local branching and merging of perturbed hidden states before generating a response. The authors say these options have complementary benefits, so a lightweight selector reads the model’s depth-wise hidden state trajectory on the prompt and chooses a computation shape. They report an 8.06% mean relative improvement on five in-distribution and 4.45% on four unseen benchmarks, with the strongest performance in 24 of 27 settings. The idea is to adapt not just how much computation a model uses, but its structure. The abstract reports benchmark results, not evidence about performance or cost in deployed systems. How well does the selector’s benchmark advantage hold up across real-world tasks and ?
Original source
cs.AI · Article
Source details
Original source date: Oct 28, 2026 (2026-10-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T03:35:40.411Z. First observed: 2026-10-30T06:45:35.314Z. Retrieved: 2026-10-31T05:17:56.337Z.
Limits
The evidence provided is the abstract, which does not describe tasks, comparison methods, or how computational cost was measured. The abstract reports benchmark evaluations, not results from deployed systems or real-world tasks. Term-library review may be needed for SelfMorph, test- scaling, depth-wise hidden state trajectory, and other technical terms in the reading.
Open question
How well does the selector’s advantage hold up across real-world tasks and ?
Source-stated
The authors introduce SelfMorph, which makes the shape of a frozen LLM’s internal computation an input-dependent inference-time choice.
We introduce SelfMorph, which makes computational shape an input-dependent inference-time decision.
Source-stated
SelfMorph selects among the original forward path, serial reuse of a layer span, and local branching and merging of perturbed hidden states.
It unifies the original forward path (BASE), serial reuse of a layer span (REPEAT), and local branching and merging of perturbed hidden states (BRANCH)
Source-stated
The authors report an 8.06% mean relative improvement on five in-distribution benchmarks and 4.45% on four unseen benchmarks.
SelfMorph achieves a 8.06% mean relative improvement on five in-distribution benchmarks and 4.45% on four unseen benchmarks.
Source-stated
The authors report that SelfMorph had the strongest performance in 24 of 27 settings.
It achieves the strongest performance in 24 of 27 settings.
Escape Meridian interpretation
The abstract reports benchmark results, but does not establish performance or cost in deployed systems.
Evaluated on three LLMs spanning 4B-27B, SelfMorph achieves a 8.06% mean relative improvement on five in-distribution benchmarks and 4.45% on four unseen benchmarks.
Source and contributor context
Wu, Junxi, Guo, Yiwen, Xia, Shu-Tao · 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.
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.
Model inference
Using a trained model to produce a prediction or response to new input.
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: 41ddf19b-ea57-4642-8d7b-6ecf0eb8cc7a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T03:35:40.411Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 128846a9a1ee387cbc904afa61d0734778eb48ac3e2238a0989c3b652943c5f1. Source text hash: d9d839713d738db8550cd101fc98e0b741560f915f88c2bdc1d682eb430ba007.
uncertain · 2026-11-13T03:35:40.411Z · anthropic/claude-opus-5.5
Publication: 41ddf19b-ea57-4642-8d7b-6ecf0eb8cc7a
attribution — pass: The reading correctly credits Junxi Wu, Yiwen Guo and Shu-Tao Xia as the authors. It keeps SelfMorph's design and results framed as the authors' introduction and reported findings rather than as established fact.
Authors:Junxi Wu, Yiwen Guo, Shu-Tao Xia
summary — uncertain: The core idea is faithfully stated: the shape of computation is chosen per input, beyond generating more tokens. But central terms are left unexplained for a general reader, including frozen LLM (weights unchanged), hidden states, layer span and in-distribution, and no tooltip glosses are shown. The reading also omits two things. First, the selector is trained to weigh task performance against computational cost, which bears directly on the closing question. Second, the tests covered three LLMs from 4B to 27B parameters. Smallest repair: briefly gloss frozen and in-distribution, and add one clause each on the cost-aware selector training and the tested model range.
We train the selector by distilling utility-based distribution from offline morphing that jointly account for task performance and computational cost.
date — pass: The originalPublished date, 2026-10-28, matches the arXiv v1 submission date. The later feed date of Oct 31 is correctly not used.
[Submitted on 28 Oct 2026]
support — pass: The stated mechanisms, the reported percentages and the 24-of-27 result are all grounded verbatim in the abstract. The interpretive claim that benchmark results do not establish deployed performance or cost is labeled as interpretation and stays within the evidence. The question is an open issue that follows from the evidence, not a leading claim.
It achieves the strongest performance in 24 of 27 settings.
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct way for a reader to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceUsing a trained model to produce a prediction or response to new input.
Definition referencearXiv 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 referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceUsing a trained model to produce a prediction or response to new input.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceUsing a trained model to produce a prediction or response to new input.
Definition referencearXiv 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 referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceUsing a trained model to produce a prediction or response to new input.
Definition reference