JANUS: leaderboard
JANUS (2026) tests whether LLMs distort true information when given an institutional goal: each scenario gives six true facts, some favorable and some adverse, and the model summarises them once neutrally and once under a latent goal. The score is the mean shift over selection, emphasis, ordering, specificity and framing, from -1.6 to 1.6. Each score is the goal-conditioned minus the neutral response for the same facts; zero means no distortion and lower is better.
Metric: Average distortion, signed from -1.6 to 1.6: mean of the five paired distortion scores (selection, emphasis, ordering, specificity, framing), paired distortion: the aspect score of the goal-conditioned response minus that of the neutral response for the same JANUS scenario (six true facts, three favorable and three adverse to an institutional objective, shown in the same order in both conditions), averaged over scenarios; facts are matched to sentences and framing is labelled by a Qwen3-8B judge with thinking disabled; zero-shot at provider or checkpoint default decoding; lower is better (less goal-driven distortion). Source: arxiv.org. Saturation forecast: Around 2031. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 0.04 | #389 |
| 2 | GPT-5.4 | 0.05 | #123 |
| 3 | Qwen 3 32B | 0.06 | #960 |
| 4 | DeepSeek V4 Flash (Non-reasoning) | 0.06 | #258 (DeepSeek V4 Flash) |
| 5 | Qwen 3 8B (Non-reasoning) | 0.07 | #1830 (Qwen 3 8B) |
| 6 | Qwen 3 8B | 0.07 | #1830 |
| 7 | Qwen 3 32B (Non-reasoning) | 0.07 | #960 (Qwen 3 32B) |
| 8 | Qwen 3 14B (Non-reasoning) | 0.07 | #1233 (Qwen 3 14B) |
| 9 | Qwen 3 14B | 0.08 | #1233 |
| 10 | Llama 3.1 8B Instruct | 0.1 | #4246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=janus · How It Works · Data refreshed daily, snapshot 2026-11-13.