mlsanityData to 14 Oct 2026

ObjexMT: Objective Extraction and Metacognitive Calibration for LLM-as-a-Judge under Multi-Turn Jailbreaks

arXiv 2508.16889 ↗First posted 23 Aug 2025cs.CLcited by 3

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

LLM-as-a-Judge (LLMaaJ) enables scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover the hidden objective of a conversation and know when that inference is reliable? Large language models degrade with irrelevant or lengthy context, and multi-turn jailbreaks can scatter goals across turns. We present ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must output a one-sentence base objective and a self-reported confidence. Accuracy is scored by semantic similarity to gold objectives, then thresholded once on 300 calibration items (τ^star = 0.66; F1@τ^star = 0.891). Metacognition is assessed with expected calibration error, Brier score, Wrong@High-Confidence (0.80 / 0.90 / 0.95), and risk--coverage curves. Across six models (gpt-4.1, claude-sonnet-4, Qwen3-235B-A22B-FP8, kimi-k2, deepseek-v3.1, gemini-2.5-flash) evaluated on SafeMTDataAttack600, SafeMTData1K, and MHJ, kimi-k2 achieves the highest objective-extraction accuracy (0.612; 95% CI [0.594, 0.630]), while claude-sonnet-4 (0.603) and deepseek-v3.1 (0.599) are statistically tied. claude-sonnet-4 offers the best selective risk and calibration (AURC 0.242; ECE 0.206; Brier 0.254). Performance varies sharply across datasets (16--82% accuracy), showing that automated obfuscation imposes challenges beyond model choice. High-confidence errors remain: [email protected] ranges from 14.9% (claude-sonnet-4) to 47.7% (Qwen3-235B-A22B-FP8). ObjexMT therefore supplies an actionable test for LLM judges: when objectives are implicit, judges often misinfer them; exposing objectives or gating decisions by confidence is advisable. All experimental data are in the Supplementary Material and at https://github.com/hyunjun1121/ObjexMTdataset.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 15 of its 22 references that are in the map. First posted on arXiv 23 Aug 2025.

Builds on

15 of 22 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference25 Jun

Other references · 14

  1. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  2. 2024A Survey on LLM-as-a-Judgecited by 1,423
  3. 2023Large Language Models Can Be Easily Distracted by Irrelevant Contextcited by 841
  4. 2023Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedbackcited by 812
  5. 2024A StrongREJECT for Empty Jailbreakscited by 385
  6. 2023A Survey of Confidence Estimation and Calibration in Large Language Modelscited by 256
  7. 2023Do Large Language Models Know What They Don't Know?cited by 243
  8. 2024GSM-Plus: A Comprehensive Benchmark for Evaluating the Robustness of LLMs as Mathematical Problem Solverscited by 158
  9. 2024LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yetcited by 151
  10. 2024LLMs know their vulnerabilities: Uncover Safety Gaps through Natural Distribution Shiftscited by 86
  11. 2024Calibrating Large Language Models Using Their Generations Onlycited by 65
  12. 2024Uncertainty in Language Models: Assessment through Rank-Calibrationcited by 41
2 more
  1. 2025M2S: Multi-turn to Single-turn jailbreak in Red Teaming for LLMscited by 8
  2. 2025Why Not Act on What You Know? Unleashing Safety Potential of LLMs via Self-Aware Guard Enhancementcited by 8

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 2; 2026: 1.

3 papers.

  1. 2025X-Teaming Evolutionary M2S: Automated Discovery of Multi-turn to Single-turn Jailbreak Templatescited by 1
  2. 2026Metacognition in LLMs: Foundations, Progress, and Opportunitiescited by 1
  3. 2025MacroBench: A Novel Testbed for Web Automation Scripts via Large Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges
  2. 2026Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data
  3. 2026JudgeStealer: Extracting LLM Judging Capabilities across Evaluation Protocols
  4. 2025Judge's Verdict: A Comprehensive Analysis of LLM Judge Capability Through Human Agreement
  5. 2026JudgeArena: A Unified Framework for Reproducible LLM-Judge Evaluation
  6. 2025Leveraging LLMs as Meta-Judges: A Multi-Agent Framework for Evaluating LLM Judgments
  7. 2024JudgeBench: A Benchmark for Evaluating LLM-based Judges
  8. 2026BabelJudge: Measuring LLM-as-a-Judge Reliability Across Languages and Agent Trajectories