Prism Hallucination Eval

prism-hallucination-eval

PRISM: Probing Reasoning, Instruction, and Source Memory in LLM Hallucinations — Wu et al. (2026) (arXiv:2604.16909, 2026)

What this evaluates

Probes LLM hallucinations across four dimensions (knowledge missing, knowledge errors, reasoning errors, and instruction-following errors) by isolating error sources through controlled, task-specific queries. It evaluates model reliability and guides optimization by measuring error rates across memory, instruction, and reasoning generation stages.

Datasets

• PRISM — total ?; splits: (unstated)

Metrics

• Accuracy — range: [0, 1]
  • Acc = (1/N) * sum(I[ŷ_i = y_i]) for closed-ended tasks, where I is the indicator function.
• LLM-Eval — range: [0, 5]
  • LLM-as-a-judge produces a scalar score s ∈ [0, 5] for open-ended tasks.
• Hallucination Rate — range: [0, 100] percent
  • S = 100 * Acc (closed) or 100 * s/5 (open); H = 100 - S.
• H-Score (primary) — range: [0, 100] percent
  • Macro-averaged hallucination rate across four dimensions: (1/4) * sum(H_d) for d in {KE, KM, RE, IFE}.

Input / output format

Input: Task-specific queries (closed-ended and open-ended) designed to probe memory, instruction, and reasoning stages, provided in a few-shot setting.

Output: Model-generated response (exact answer for closed-ended tasks, free-form text for open-ended tasks).

Scoring recipe

def compute_prism_metrics(predictions, golds, is_open_ended, llm_scores=None):
    if is_open_ended:
        acc = [s / 5.0 for s in llm_scores]
    else:
        acc = [1.0 if p == g else 0.0 for p, g in zip(predictions, golds)]
    unified_scores = [100.0 * a for a in acc]
    h_rates = [100.0 - s for s in unified_scores]
    return {
        'accuracy': sum(acc) / len(acc),
        'hallucination_rate': sum(h_rates) / len(h_rates),
        'h_score': sum(h_rates) / len(h_rates)
    }

Common pitfalls

• Confusing the unified score S with the hallucination rate H, as they are exact complements (H = 100 - S).
• Using a single generation temperature/top-p across all tasks; the paper specifies different optimal parameters for closed-ended vs. open-ended tasks and across dimensions (e.g., RE requires temp=0.4, top-p=0.95 for stability).
• Assuming step-by-step explanations (CoT) reliably mitigate hallucinations; the paper notes they often fail to yield consistent gains and may be unfaithful to internal model states.

Evidence (verbatim from paper)

We employ distinct metrics for each subtask to enable a hallucination comparison. Accuracy: For closed-ended tasks, we employ standard Accuracy. LLM-Eval: For open-ended tasks, we adopt a LLM evaluator following LLM-Eval, which produces a scalar score s∈[0,5]. Hallucination Rate: We first map all task metrics to a unified percentage score S∈[0,100] and define the hallucination rate as its complement: H=100−S. H-Score: Let Hd denote the macro-averaged hallucination rate for each dimension d∈D={KE,KM,RE,IFE}. We define H-Score=(1/4)∑d∈D Hd.

Citation

@misc{wu2026prism,
  title={PRISM: Probing Reasoning, Instruction, and Source Memory in LLM Hallucinations},
  author={Wu et al. (2026)},
  year={2026},
  note={arXiv:2604.16909}
}

• arXiv: 2604.16909