Jao, Jonathan F., Lampinen, Andrew Kyle · Oct 28, 2026
Jao and Lampinen propose Value Probes for language-model red teaming
Markdown · JSON · Original source
In their paper, Jonathan F. Jao and Andrew Kyle Lampinen argue that automated audits can search more effectively by treating possible conversation paths as a Markov decision process and using beam search. A red-team audit probes a model for safety-relevant failures; beam search keeps several promising conversation paths in play rather than following only one auditor model’s choices. Their method, Value Probes, uses linear probing of the target model’s activations to estimate discounted future rewards: judge scores that represent the target behavior under audit. The authors report that, across 18 target behaviors and three model scales, this approach raises peak audit scores by roughly 50% against token-matched standard audits. They also report that Value Probes can be trained with off-policy data from smaller models, and offer a comparably performing, more efficient alternative to more expensive language-model follow-ups. The authors’ case is that better search can help auditors find problematic behavior without relying on the auditor model alone. The reported results concern audit scores, not the frequency of failures in deployment or whether models become safer. The abstract does not specify the target behaviors, model details, or how the reported gains vary across them. How well do the audit-score gains hold across different target behaviors and model scales
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: 2026-11-10T19:18:41.794Z. First observed: 2026-10-30T06:42:51.314Z. Retrieved: 2026-10-31T10:02:24.955Z.
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 available evidence is the abstract, which does not identify the target behaviors or provide model details, per-behavior results, or the spread of results across model scales. The reported improvement is in peak audit scores; the abstract does not establish that deployment failures become less frequent or that models become safer. The abstract does not provide enough detail to assess how the off-policy data or efficiency comparison was evaluated.
Open question
How well do the audit-score gains hold across different target behaviors and model scales?
Source-stated
Jao and Lampinen frame automated red-team auditing as searching conversation pathways, and propose beam search instead of relying on the auditor model alone.
we propose adopting beam search for the audit rather than relying on the auditor alone.
Source-stated
The authors introduce Value Probes, which use linear probing of target-model activations to predict discounted future rewards and guide beam-search audits.
this value function can be estimated by linear probing from the target model's activations to predict discounted future rewards - a method we term "Value Probes."
Source-stated
The authors report roughly 50% higher peak audit scores than token-matched standard approaches across 18 target behaviors and three model scales.
Across 18 target behaviors and 3 scales of models, our approach raises peak audit scores by roughly 50% compared to (token-matched) standard auditing approaches.
Source-stated
The authors say Value Probes can be fit using off-policy data from smaller models and can offer a comparably performing, more efficient alternative to more expensive language-model follow-ups.
Value Probes can be fit using off-policy data from smaller models. We also ablate individual components, and show that our method offers a comparably-performant and more efficient alternative to more expensive language model follow-ups.
Source and contributor context
Jao, Jonathan F., Lampinen, Andrew Kyle · 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.
Red-teaming
Deliberately attacking or tricking a system to find its weaknesses before others do.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d08e4bfd-9589-4035-8200-453f5b7b6e7c.
Prepared by server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna); grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T19:18:41.794Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: fcf025461af24afb582dd7d2da30a3436a003d13e68be9f0ed358a13794edaf1. Source text hash: e7783b4f2129e6e02964860a5246dc36423fd1bf1cc518be2b07cf57e783ca3a.
Revision and source-check history
Version 1 · 2026-11-10T19:18:41.794Z
Jao and Lampinen propose Value Probes for language-model red teaming
In their paper, Jonathan F. Jao and Andrew Kyle Lampinen argue that automated audits can search more effectively by treating possible conversation paths as a Markov decision process and using beam search. A red-team audit probes a model for safety-relevant failures; beam search keeps several promising conversation paths in play rather than following only one auditor model’s choices. Their method, Value Probes, uses linear probing of the target model’s activations to estimate discounted future rewards: judge scores that represent the target behavior under audit. The authors report that, across 18 target behaviors and three model scales, this approach raises peak audit scores by roughly 50% against token-matched standard audits. They also report that Value Probes can be trained with off-policy data from smaller models, and offer a comparably performing, more efficient alternative to more expensive language-model follow-ups. The authors’ case is that better search can help auditors find problematic behavior without relying on the auditor model alone. The reported results concern audit scores, not the frequency of failures in deployment or whether models become safer. The abstract does not specify the target behaviors, model details, or how the reported gains vary across them. How well do the audit-score gains hold across different target behaviors and model scales
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: 2026-11-10T19:18:41.794Z. First observed: 2026-10-30T06:42:51.314Z. Retrieved: 2026-10-31T10:02:24.955Z.
Limits
The available evidence is the abstract, which does not identify the target behaviors or provide model details, per-behavior results, or the spread of results across model scales. The reported improvement is in peak audit scores; the abstract does not establish that deployment failures become less frequent or that models become safer. The abstract does not provide enough detail to assess how the off-policy data or efficiency comparison was evaluated.
Open question
How well do the audit-score gains hold across different target behaviors and model scales?
Source-stated
Jao and Lampinen frame automated red-team auditing as searching conversation pathways, and propose beam search instead of relying on the auditor model alone.
we propose adopting beam search for the audit rather than relying on the auditor alone.
Source-stated
The authors introduce Value Probes, which use linear probing of target-model activations to predict discounted future rewards and guide beam-search audits.
this value function can be estimated by linear probing from the target model's activations to predict discounted future rewards - a method we term "Value Probes."
Source-stated
The authors report roughly 50% higher peak audit scores than token-matched standard approaches across 18 target behaviors and three model scales.
Across 18 target behaviors and 3 scales of models, our approach raises peak audit scores by roughly 50% compared to (token-matched) standard auditing approaches.
Source-stated
The authors say Value Probes can be fit using off-policy data from smaller models and can offer a comparably performing, more efficient alternative to more expensive language-model follow-ups.
Value Probes can be fit using off-policy data from smaller models. We also ablate individual components, and show that our method offers a comparably-performant and more efficient alternative to more expensive language model follow-ups.
Source and contributor context
Jao, Jonathan F., Lampinen, Andrew Kyle · 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.
Red-teaming
Deliberately attacking or tricking a system to find its weaknesses before others do.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d08e4bfd-9589-4035-8200-453f5b7b6e7c.
Prepared by server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna); grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T19:18:41.794Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: fcf025461af24afb582dd7d2da30a3436a003d13e68be9f0ed358a13794edaf1. Source text hash: e7783b4f2129e6e02964860a5246dc36423fd1bf1cc518be2b07cf57e783ca3a.
uncertain · 2026-11-10T19:18:41.794Z · anthropic/claude-opus-5.5
Publication: d08e4bfd-9589-4035-8200-453f5b7b6e7c
attribution — pass: The paper is correctly attributed to Jonathan F. Jao and Andrew Kyle Lampinen on arXiv. Results are framed as the authors' reports ('the authors report'), not as established fact.
Authors:Jonathan F. Jao, Andrew Kyle Lampinen
summary — uncertain: The content is faithful, but the summary ends with a stray, unpunctuated copy of the question ('How well do the audit-score gains hold...'). Smallest repair: delete that fragment. Several central terms are not explained for general readers: Markov decision process, activations, off-policy data, and token-matched. Smallest repair: add one short plain-language gloss for each, or term-library tooltips. The title and the question fit the source.
date — pass: originalPublished 2026-10-28 matches the v1 submission date. The later feed date of 2026-10-31 is correctly not used.
[Submitted on 28 Oct 2026]
support — pass: All four claims are directly stated in the abstract: the MDP framing with beam search, Value Probes via linear probing, roughly 50% higher peak audit scores across 18 behaviors and 3 scales, and off-policy fitting with an efficiency comparison. The caveats correctly avoid turning audit scores into a claim of real-world safety. The question is not leading.
Across 18 target behaviors and 3 scales of models, our approach raises peak audit scores by roughly 50% compared to (token-matched) standard auditing approaches.
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct way 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.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivDeliberately attacking or tricking a system to find its weaknesses before others do.
Definition referencearXiv 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 arXivDeliberately attacking or tricking a system to find its weaknesses before others do.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv