CIPHER Cross-Record Inference - FIR (Few-Shot): leaderboard

Metric: Harmonic answer score (%; per question the harmonic mean of token overlap with the reference and Gemini 2.5 Flash semantic acceptance, zero when rejected, averaged; First Information Reports (Indian police records) with typed placeholders from post-hoc masking; the full support-controlled context of 39-42 records including distractors; two-example few-shot prompting; greedy decoding). Source: arxiv.org. Saturation forecast: Around December 2026. 5 models tracked.

Top models

#ModelScore
1Gemini 2.0 Flash42.4
2Qwen 3 Next 80B A3B37.3
3GPT-OSS-120B34.7
4Llama 3.3 70B Instruct32.8
5GPT-OSS-20B22.8

Interactive version: theaggregate.ai/benchmark?slug=cipher-cross-record-inference-fir-few-shot · How It Works · Data refreshed daily, snapshot 2026-09-26.