Van Deventer, Hugh, Troitskii, Dmitrii, Ruis, Laura · Oct 29, 2026
The First Few Tokens Matter: Understanding Chain-of-Thought Controllability
Markdown · JSON · Original source
Hugh Van Deventer, Dmitrii Troitskii and Laura Ruis argue that models’ difficulty controlling what appears in their (CoT) traces may stem from an entropy collapse in the first few tokens after the opening thinking tag. CoT monitoring uses those reasoning traces to oversee increasingly capable agents, the authors write, but this oversight depends in part on models being limited in their ability to control what they verbalize. The authors report two inference- interventions on initial tokens: prefilling, where even a single token can substantially increase controllability, and banning frequent opening tokens alongside in-context examples. Across models, the latter improves average compliance by 34-41 percentage points. The reported gains apply mainly to addition and style controllability; suppression controllability has more varied results and often shows a different relationship with monitorability. The authors argue that these patterns call for evaluations that distinguish difficulty initiating a particular opening from broader difficulty altering reasoning content, and that report control tasks and categories separately. The abstract does not establish how well these interventions or evaluation patterns hold beyond the reported experiments.
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 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T00:26:00.801Z. First observed: 2026-11-02T05:04:37.956Z. Retrieved: 2026-11-02T07:50:45.248Z.
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 abstract and page metadata, not the paper’s full methods or results; it does not provide the experimental setup or enough detail to assess how broadly the reported patterns apply.
Open question
How do the reported relationships between controllability and monitorability vary across the paper’s evaluation settings?
Source-stated
The authors link limits on models’ ability to control what they verbalize in Chain-of-thought traces to an entropy collapse in the first few tokens after the opening thinking tag.
We investigate the source of this limitation and link it to an entropy collapse in the first few tokens following the opening thinking tag.
Source-stated
The authors report two inference-time interventions on initial tokens: prefilling and banning frequent opening tokens combined with in-context examples.
two inference-time interventions on initial tokens: prefilling, where even a single token can substantially increase controllability, and banning frequent opening tokens combined with in-context examples
Source-stated
Banning frequent opening tokens combined with in-context examples improves average compliance by 34-41 percentage points across models, according to the abstract.
banning frequent opening tokens combined with in-context examples, which improves average compliance by 34-41 percentage points across models
Source-stated
The abstract says the interventions mainly recover addition and style controllability, while suppression controllability exhibits more varied responses.
These interventions recover mainly addition and style controllability (adding or transforming information), while suppression controllability (avoiding or removing information) exhibits more varied responses
Source-stated
The authors say the relationship between controllability and monitorability depends on the evaluation, with suppression often showing negative associations.
addition and style controllability more consistently positively correlate with monitorability, while suppression often shows negative associations
Source-stated
The authors recommend evaluations that distinguish difficulty initiating a specific opening from general limits on altering reasoning content, and report control tasks and categories separately.
Our findings motivate the development of controllability evaluations that distinguish difficulty initiating a specific opening from general limitations of the ability to alter reasoning content, alongside separate reporting of control tasks and categories.
Source and contributor context
Van Deventer, Hugh, Troitskii, Dmitrii, Ruis, Laura · 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.
Chain of thought
The step-by-step working a model writes out before giving its final answer.
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: ecc8d497-354b-4ad2-8a14-f8289ece203a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T00:26:00.801Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: ba4fd8b8cc5de4beeec631ef8e09e63f9ebade86da17a79de16f47392d84f711. Source text hash: 418dd127fae0f4c1d35242ae089f5cb147cf136ee72655c7c8655195ad50371b.
Revision and source-check history
Version 1 · 2026-11-11T00:26:00.801Z
The First Few Tokens Matter: Understanding Chain-of-Thought Controllability
Hugh Van Deventer, Dmitrii Troitskii and Laura Ruis argue that models’ difficulty controlling what appears in their (CoT) traces may stem from an entropy collapse in the first few tokens after the opening thinking tag. CoT monitoring uses those reasoning traces to oversee increasingly capable agents, the authors write, but this oversight depends in part on models being limited in their ability to control what they verbalize. The authors report two inference- interventions on initial tokens: prefilling, where even a single token can substantially increase controllability, and banning frequent opening tokens alongside in-context examples. Across models, the latter improves average compliance by 34-41 percentage points. The reported gains apply mainly to addition and style controllability; suppression controllability has more varied results and often shows a different relationship with monitorability. The authors argue that these patterns call for evaluations that distinguish difficulty initiating a particular opening from broader difficulty altering reasoning content, and that report control tasks and categories separately. The abstract does not establish how well these interventions or evaluation patterns hold beyond the reported experiments.
Original source
cs.AI · Article
Source details
Original source date: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T00:26:00.801Z. First observed: 2026-11-02T05:04:37.956Z. Retrieved: 2026-11-02T07:50:45.248Z.
Limits
The supplied evidence is the abstract and page metadata, not the paper’s full methods or results; it does not provide the experimental setup or enough detail to assess how broadly the reported patterns apply.
Open question
How do the reported relationships between controllability and monitorability vary across the paper’s evaluation settings?
Source-stated
The authors link limits on models’ ability to control what they verbalize in Chain-of-thought traces to an entropy collapse in the first few tokens after the opening thinking tag.
We investigate the source of this limitation and link it to an entropy collapse in the first few tokens following the opening thinking tag.
Source-stated
The authors report two inference-time interventions on initial tokens: prefilling and banning frequent opening tokens combined with in-context examples.
two inference-time interventions on initial tokens: prefilling, where even a single token can substantially increase controllability, and banning frequent opening tokens combined with in-context examples
Source-stated
Banning frequent opening tokens combined with in-context examples improves average compliance by 34-41 percentage points across models, according to the abstract.
banning frequent opening tokens combined with in-context examples, which improves average compliance by 34-41 percentage points across models
Source-stated
The abstract says the interventions mainly recover addition and style controllability, while suppression controllability exhibits more varied responses.
These interventions recover mainly addition and style controllability (adding or transforming information), while suppression controllability (avoiding or removing information) exhibits more varied responses
Source-stated
The authors say the relationship between controllability and monitorability depends on the evaluation, with suppression often showing negative associations.
addition and style controllability more consistently positively correlate with monitorability, while suppression often shows negative associations
Source-stated
The authors recommend evaluations that distinguish difficulty initiating a specific opening from general limits on altering reasoning content, and report control tasks and categories separately.
Our findings motivate the development of controllability evaluations that distinguish difficulty initiating a specific opening from general limitations of the ability to alter reasoning content, alongside separate reporting of control tasks and categories.
Source and contributor context
Van Deventer, Hugh, Troitskii, Dmitrii, Ruis, Laura · 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.
Chain of thought
The step-by-step working a model writes out before giving its final answer.
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: ecc8d497-354b-4ad2-8a14-f8289ece203a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T00:26:00.801Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: ba4fd8b8cc5de4beeec631ef8e09e63f9ebade86da17a79de16f47392d84f711. Source text hash: 418dd127fae0f4c1d35242ae089f5cb147cf136ee72655c7c8655195ad50371b.
uncertain · 2026-11-11T00:26:00.801Z · anthropic/claude-opus-5.5
Publication: ecc8d497-354b-4ad2-8a14-f8289ece203a
attribution — pass: The reading correctly credits Hugh Van Deventer, Dmitrii Troitskii and Laura Ruis as the arXiv authors. It keeps findings and recommendations attributed to them with phrasing such as 'the authors report,' 'argue' and 'according to the abstract.'
Authors:Hugh Van Deventer, Dmitrii Troitskii, Laura Ruis
summary — uncertain: The summary is faithful but may not stand alone for a general reader. First, central terms like 'entropy collapse', 'thinking tag', 'prefilling' and 'monitorability' are unexplained; they need term-library tooltips. Second, the draft drops the abstract's own glosses: addition/style means adding or transforming information, and suppression means avoiding or removing it. Third, the key stake is left implicit: monitoring works partly because models cannot control their reasoning, and simple interventions restore that control. Fourth, it omits that the two control groups behave as separate capabilities. Fifth, 'a different relationship' blurs the source's 'often shows negative associations.'
suppression controllability (avoiding or removing information) exhibits more varied responses, suggesting distinct underlying mechanisms
date — pass: originalPublished is 2026-10-29. This matches the arXiv v1 submission date, not the later feed date of 2026-11-02.
[Submitted on 29 Oct 2026]
support — pass: Every claim matches its quoted abstract text, including the 34-41 percentage point figure and the evaluation recommendation. The draft notes that only the abstract was available and does not turn the reported gains into general reliability. The question follows from the abstract's statement that the relationship depends on the evaluation.
we find that the relationship with monitorability depends on the evaluation
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to 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.
The step-by-step working a model writes out before giving its final answer.
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 arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe step-by-step working a model writes out before giving its final answer.
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 arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv