Skip to content
OpenTrain AIFor AI Companies

HFEPX · Eval paper review

Do Thinking Tokens Help with Safety?

Narutatsu Ri, Abhishek Panigrahi, Sanjeev Arora

Published

Jun 23, 2026

Citations

0

Trust level

Moderate

Usefulness score

65/100 (Medium)

Extraction confidence

70% (Moderate)

Derived from extracted protocol signals and abstract evidence.

Rater population

Not reported

Signals refreshed

Jun 23, 2026

Open arXiv Find implementation Shortlist (0) RSS feed

Tags

Automatic Metrics Critique Edit General

Should you rely on this paper?

This paper has useful evaluation signal, but protocol completeness is partial; pair it with related papers before deciding implementation strategy.

Use this for comparison and orientation, not as your only source.

Best use

Secondary protocol comparison source

Use if you need

A secondary eval reference to pair with stronger protocol papers.

What to verify

Validate the evaluation procedure and quality controls in the full paper before operational use.

Main weakness

No major weakness surfaced.

Human feedback signal
Detected
From extracted signals
Evaluation signal
Detected
Eval setup described
Usefulness for eval research
65/100
Moderate-confidence candidate

Useful as a secondary reference; validate protocol details against neighboring papers.

If you are doing eval pipeline work, start here

Human Eval Papers LLM-as-Judge Papers Pairwise Preference Papers Tool-Use Evaluation Search

Abstract

Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts. It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles. We present evidence that this intuition is not always correct. Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation ($0.84$-$0.95$ AUROC and $\sim88\%$ balanced accuracy for predicting refusal/compliance) before any visible thinking. The thinking process turns out to be more akin to prefix completion than to deliberative revision, with the final outcome rarely changing after the first $\sim20\%$ of thinking, despite giving the appearance of deliberation at the text level ($\sim74\%$ of text-level deliberations occur when the response distribution is already locked to one refusal/compliance side). We also find that existing inference-time and training-based safety interventions, despite being motivated by the goal of inducing deliberation, largely shift model behavior toward over-refusal while suppressing already-scarce deliberation signals. Our results suggest that safety behavior in current reasoning models is much less deliberative than commonly assumed, and highlight the need for methods that induce real safety deliberation.

What we could verify

These are the protocol signals we could actually recover from the available paper metadata. Use them to decide whether this paper is worth deeper reading.

Human Feedback Types

strong

Critique Edit

Directly usable for protocol triage.

"Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts."

Evaluation Modes

strong

Automatic Metrics

Includes extracted eval setup.

"Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts."

Quality Controls

missing

Not reported

No explicit QC controls found.

"Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts."

Benchmarks / Datasets

missing

Not extracted

No benchmark anchors detected.

"Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts."

Reported Metrics

strong

Accuracy, Auroc

Useful for evaluation criteria comparison.

"Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation ($0.84$-$0.95$ AUROC and $\sim88\%$ balanced accuracy for predicting refusal/compliance) before any visible thinking."

Benchmarks and datasets

No benchmark or dataset names were extracted from the available abstract.

Reported metrics

accuracyauroc
Human feedback details
Uses human feedback
Yes
Feedback types
Critique Edit
Rater population
Not reported
Expertise required
General
Evaluation details
Evaluation modes
Automatic Metrics
Agentic eval
None
Quality controls
Not reported
Evidence quality
Moderate
Use this page as
Secondary protocol comparison source

Research brief

Metadata summary

Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts.

Based on abstract + metadata only. Check the source paper before making high-confidence protocol decisions.

Key takeaways

  • Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts.
  • It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles.
  • We present evidence that this intuition is not always correct.

Researcher actions

  • Compare this paper against nearby papers in the same arXiv category before using it for protocol decisions.
  • Validate inferred eval signals (Automatic metrics) against the full paper.
  • Use related-paper links to find stronger protocol-specific references.

Caveats

  • Generated from abstract + metadata only; no PDF parsing.
  • Signals below are heuristic and may miss details reported outside the abstract.

Recommended queries

Accuracy

Contribution summary

  • Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts.
  • It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles.
  • We present evidence that this intuition is not always correct.

Why it matters for eval

  • Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts.
  • It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles.

Researcher checklist

  • Human feedback protocol is explicit

    Detected: Critique Edit

  • Evaluation mode is explicit

    Detected: Automatic Metrics

  • Quality control reporting appears

    No calibration/adjudication/IAA control explicitly detected.

  • Benchmark or dataset anchors are present

    No benchmark/dataset anchor extracted from abstract.

  • Metric reporting is present

    Detected: accuracy, auroc