跳到正文
DEV Community· AI OpenFree·· 11 天前AI 評分49

JGENai 與 VIDRAFT 推出評估 LLM 自我錯誤偵測能力的基準測試

Metacognition Leaderboard for LLMs: JGENai VIDRAFT Launch a Benchmark That Measures Self-Error Detection, Not Just Accuracy

AI 導讀

JGENai 與 VIDRAFT 聯合發布元認知基準測試、24 個模型的公開排行榜及 11 款模型專用元認知適配器,用來評估模型能否察覺自身錯誤。基準包含 400 道陷阱題,分別評估選擇題準確率及開放式自我錯誤判斷能力;JGOS-31B-Citizen 排名 24 個模型之首,答對 398 題,陷阱題迴避率達 99.5%。

正文

Metacognition Leaderboard for LLMs: JGENai × VIDRAFT Launch a Benchmark That Measures Self-Error Detection, Not Just Accuracy

TL;DR: JGENai and VIDRAFT have jointly released a metacognition benchmark, a 24-model public leaderboard, and 11 model-specific metacognition adapters designed to measure whether an AI can detect its own mistakes — not just whether it gets answers right. The adapters attach to existing models without modifying weights or architecture, making them drop-in reliability probes for high-stakes applications. Developers can explore the leaderboard, benchmark dataset, and adapters on Hugging Face.

What it is

JGENai (CEO Sunyoung Choi) and VIDRAFT have co-developed and publicly released three artifacts:

  • A metacognition benchmark — 400 adversarial "trap" questions designed so the superficially plausible answer is wrong, testing whether a model can catch its own likely errors rather than just pattern-match to a confident-sounding output.
  • A 24-model public leaderboard — ranking current Korean and multilingual LLMs on both multiple-choice accuracy and open-ended self-error-detection capability.
  • 11 model-specific metacognition adapters — lightweight add-ons that sit beside a target model and assess the risk that any given response is incorrect, without touching the base model's weights.

The core thesis: the dangerous moment for an AI is not when it's wrong, but when it doesn't know it's wrong. In domains like medicine, law, finance, security, and scientific research — where error costs are high — self-error detection is framed as a primary reliability signal, not an afterthought.

How it works

Benchmark design

Each of the 400 benchmark problems is constructed so that one answer appears strongly plausible at first read, while the actual correct answer requires the model to pause and re-examine its own reasoning. Evaluation is split across two axes:

  1. Multiple-choice accuracy — standard answer correctness on trap questions.
  2. Open-ended self-assessment — after generating a free-form answer, can the model correctly judge whether that answer is likely wrong?

This dual-axis design exposes a gap that single-accuracy leaderboards miss: a model can score well on axis 1 while being nearly blind on axis 2.

Metacognition adapters

The adapters are designed around a non-invasive attachment pattern:

  • They do not modify the base model's architecture or trained weights.
  • They analyze internal signals produced during the model's answer-generation process to compute a risk score indicating how likely the response is to be incorrect.
  • The output is a separate risk signal — a flag or score that downstream systems, human reviewers, or orchestration layers can consume independently of the answer itself.

This architecture means you can wrap an existing deployed model with a metacognition adapter without retraining, fine-tuning, or touching inference infrastructure beyond adding the adapter layer.

Benchmarks & results

From the publicly reported leaderboard results:

  • JGOS-31B-Citizen ranked #1 out of 24 models evaluated. It answered 398 out of 400 trap questions correctly, achieving a 99.5% trap-avoidance rate. JGOS-31B-Citizen is also the current top-ranked model on the K-AI Leaderboard operated by Korea's Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA).
  • Multiple models from the Darwin family (Darwin-28B, Darwin-31B, Darwin-35B) — all developed in Korea — placed in the upper tier of the leaderboard.
  • Critically, models that performed strongly on multiple-choice accuracy still showed limited capability on open-ended self-assessment. This confirms that raw accuracy and metacognitive awareness are empirically distinct competencies — a model can be highly accurate yet unable to flag its own uncertain or incorrect outputs.

How to try it

The benchmark dataset, the 24-model leaderboard, and all 11 metacognition adapters are publicly available on Hugging Face. The source article does not specify exact repository slugs, so navigate directly to the JGENai organization page on Hugging Face to find the released artifacts.

# Browse available artifacts via the Hugging Face CLI
huggingface-cli scan-cache

# Or search the JGENai org directly in your browser:
# https://huggingface.co/JGENai

No GitHub repository URL, API endpoint, or additional access mechanism was mentioned in the source article at time of publication. Check the Hugging Face model cards for adapter usage instructions specific to each of the 11 supported models.

FAQ

Q: Does applying a metacognition adapter require retraining or fine-tuning my existing model?
A: No. The adapters are explicitly designed to leave the base model's weights and architecture unchanged. They attach externally and analyze generation-time signals to produce a separate risk score.

Q: Is the benchmark only relevant for Korean-language models?
A: The benchmark was developed with a Korean-language focus — the top-performing models are Korean LLMs — but the conceptual framework (trap questions + dual-axis evaluation) is language-agnostic. Whether English or multilingual models are included in the current 24-model leaderboard is not specified in the source article.

Q: What does the risk signal from the adapter actually look like in practice?
A: The source describes it as a computed probability that a given response is incorrect, derived from internal signals during generation. Exact output format (scalar score, binary flag, confidence interval) is not specified in the public reporting — check the Hugging Face model cards for implementation details.

Q: Why does it matter that accuracy and self-error detection are separate skills?
A: Because a high-accuracy model that cannot flag its own mistakes will confidently surface wrong answers in production with no warning. In regulated or safety-critical domains, that silent failure mode is often more dangerous than a lower-accuracy model that correctly signals uncertainty.


Originally reported by AI이코노미뉴스 (2026-07-02) — source article.

來源:DEV Community · dev.to