E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning
The paper tests whether financial chart VLMs can carry evidence all the way to an action, not just avoid unsupported claims.
E2A-Bench contains 969 queries built from 323 HS300 constituents, with OHLCV-derived evidence anchors. The authors evaluate 20 VLMs on grounding, reasoning-action consistency, confidence calibration, and directional coverage. They report that scalar hallucination scores can miss failures such as low directional coverage, coverage collapse after oracle-aided verification, and BUY-skew amplification after financial fine-tuning. HF Daily Papers' note
E2A-Bench contains 969 queries built from 323 HS300 constituents, with OHLCV-derived evidence anchors. The authors evaluate 20 VLMs on grounding, reasoning-action consistency, confidence calibration, and directional coverage. They report that scalar hallucination scores can miss failures such as low directional coverage, coverage collapse after oracle-aided verification, and BUY-skew amplification after financial fine-tuning. HF Daily Papers' note
score 4