Skip to content
← Back to explorer

Stealth Fine-Tuning: Efficiently Breaking Alignment in RVLMs Using Self-Generated CoT

Le Yu, Zhengyue Zhao, Yawen Zheng, Yunhao Liu · Nov 18, 2025 · Citations: 0

Automatic Metrics
Open arXiv Find Implementation RSS feed Shortlist (0)

Data freshness

Extraction: Fresh

Check recency before relying on this page for active eval decisions. Use stale pages as context and verify against current hub results.

Metadata refreshed

Mar 7, 2026, 9:21 AM

Recent

Extraction refreshed

Mar 14, 2026, 3:26 AM

Fresh

Extraction source

Persisted extraction

Confidence 0.45

Abstract

Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces. In this work, we find that the safety alignment of RVLMs can be easily broken through a novel attack method termed \textbf{Stealth Fine-Tuning}. Our method elicits harmful reasoning traces through \textbf{segment-level interference} and reuses the self-generated outputs as supervised fine-tuning data. To facilitate this, we introduce a \textbf{turn-based weighted} loss that minimizes distribution shift. In our experiment, with only 499 samples and under 3 hours on a single A100 (QLoRA), Stealth Fine-Tuning outperforms IDEATOR by 38.66\% ASR while preserving general reasoning ability, as the tuned model retains the original representation distribution. Experiments on AdvBench and several general benchmarks demonstrate that Stealth Fine-Tuning is a low-cost and highly effective way to bypass alignment defenses. \textcolor{red}{\textbf{Disclaimer: This paper contains content that may be disturbing or offensive.}}

Low-signal caution for protocol decisions

Use this page for context, then validate protocol choices against stronger HFEPX references before implementation decisions.

  • Extraction flags indicate low-signal or possible false-positive protocol mapping.
  • Extraction confidence is 0.45 (below strong-reference threshold).
Human Eval Hub LLM-as-Judge Hub Pairwise Preference Hub

HFEPX Relevance Assessment

This paper is adjacent to HFEPX scope and is best used for background context, not as a primary protocol reference.

Best use

Background context only

Use if you need

A benchmark-and-metrics comparison anchor.

Main weakness

Extraction flags indicate low-signal or possible false-positive protocol mapping.

Trust level

Low

Eval-Fit Score

5/100 • Low

Treat as adjacent context, not a core eval-method reference.

Human Feedback Signal

Not explicit in abstract metadata

Evaluation Signal

Detected

HFEPX Fit

Adjacent candidate

Extraction confidence: Low

If you are doing eval pipeline work, start here:

Human Eval Hub LLM-as-Judge Hub Pairwise Preference Hub Tool-Use Eval Hub

Field Provenance & Confidence

Each key protocol field shows extraction state, confidence band, and data source so you can decide whether to trust it directly or validate from full text.

Human Feedback Types

missing

None explicit

Confidence: Low Source: Persisted extraction missing

No explicit feedback protocol extracted.

Evidence snippet: Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces.

Evaluation Modes

partial

Automatic Metrics

Confidence: Low Source: Persisted extraction evidenced

Includes extracted eval setup.

Evidence snippet: Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces.

Quality Controls

missing

Not reported

Confidence: Low Source: Persisted extraction missing

No explicit QC controls found.

Evidence snippet: Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces.

Benchmarks / Datasets

partial

AdvBench

Confidence: Low Source: Persisted extraction evidenced

Useful for quick benchmark comparison.

Evidence snippet: Experiments on AdvBench and several general benchmarks demonstrate that Stealth Fine-Tuning is a low-cost and highly effective way to bypass alignment defenses.

Reported Metrics

partial

Cost, Jailbreak success rate

Confidence: Low Source: Persisted extraction evidenced

Useful for evaluation criteria comparison.

Evidence snippet: Experiments on AdvBench and several general benchmarks demonstrate that Stealth Fine-Tuning is a low-cost and highly effective way to bypass alignment defenses.

Rater Population

missing

Unknown

Confidence: Low Source: Persisted extraction missing

Rater source not explicitly reported.

Evidence snippet: Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces.

Human Data Lens

  • Uses human feedback: No
  • Feedback types: None
  • Rater population: Unknown
  • Unit of annotation: Unknown
  • Expertise required: General
  • Extraction source: Persisted extraction

Evaluation Lens

  • Evaluation modes: Automatic Metrics
  • Agentic eval: None
  • Quality controls: Not reported
  • Confidence: 0.45
  • Flags: low_signal, possible_false_positive

Protocol And Measurement Signals

Benchmarks / Datasets

AdvBench

Reported Metrics

costjailbreak success rate

Research Brief

Deterministic synthesis

Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces. HFEPX signals include Automatic Metrics with confidence 0.45. Updated from current HFEPX corpus.

Generated Mar 14, 2026, 3:26 AM · Grounded in abstract + metadata only

Key Takeaways

  • Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack…
  • In this work, we find that the safety alignment of RVLMs can be easily broken through a novel attack method termed Stealth Fine-Tuning.

Researcher Actions

  • Treat this as method context, then pivot to protocol-specific HFEPX hubs.
  • Cross-check benchmark overlap: AdvBench.
  • Validate metric comparability (cost, jailbreak success rate).

Caveats

  • Generated from title, abstract, and extracted metadata only; full-paper implementation details are not parsed.
  • Low-signal flag detected: protocol relevance may be indirect.

Research Summary

Contribution Summary

  • Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces.
  • In this work, we find that the safety alignment of RVLMs can be easily broken through a novel attack method termed Stealth Fine-Tuning.
  • To facilitate this, we introduce a turn-based weighted loss that minimizes distribution shift.

Why It Matters For Eval

  • Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces.
  • In this work, we find that the safety alignment of RVLMs can be easily broken through a novel attack method termed Stealth Fine-Tuning.

Researcher Checklist

  • Gap: Human feedback protocol is explicit

    No explicit human feedback protocol detected.

  • Pass: Evaluation mode is explicit

    Detected: Automatic Metrics

  • Gap: Quality control reporting appears

    No calibration/adjudication/IAA control explicitly detected.

  • Pass: Benchmark or dataset anchors are present

    Detected: AdvBench

  • Pass: Metric reporting is present

    Detected: cost, jailbreak success rate

Category-Adjacent Papers (Broader Context)

These papers are nearby in arXiv category and useful for broader context, but not necessarily protocol-matched to this paper.

Need human evaluators for your AI research? Scale annotation with expert AI Trainers.

Post a Job Get a Quote