Context Conflict Merge Eval

context-conflict-merge-eval

Blinded by Generated Contexts: How Language Models Merge Generated and Retrieved Contexts When Knowledge Conflicts? — Hexiang Tan et al. (2024) (arXiv:2401.11911, 2024)

What this evaluates

Evaluates how language models merge conflicting generated and retrieved contexts in open-domain QA. It probes whether models exhibit a systematic bias toward generated contexts over retrieved ones when only one context contains the correct answer.

Datasets

• NQ-CC — total ?; splits: test (-1), dev (-1); repo https://github.com/Tan-Hexiang/RetrieveOrGenerated
• TQA-CC — total ?; splits: test (-1), dev (-1); repo https://github.com/Tan-Hexiang/RetrieveOrGenerated

Metrics

• DiffGR (primary) — range: [-1, 1]
  • DiffGR = (ρ_gen - ρ_ret) / (ρ_gen + ρ_ret), where ρ_gen = avg(exact_match(model_answer, generated_only_answer)) and ρ_ret = avg(exact_match(model_answer, retrieved_only_answer)). Ranges from [-1, 1].

Input / output format

Input: Question q, retrieved context d_1^γ, and generated context d_1^ϱ provided together as input to the LLM.

Output: A single generated answer string a_φ.

Scoring recipe

def compute_diffgr(predictions, gold_gen, gold_ret):
    em_gen = [1 if p == g else 0 for p, g in zip(predictions, gold_gen)]
    em_ret = [1 if p == g else 0 for p, g in zip(predictions, gold_ret)]
    rho_gen = sum(em_gen) / len(em_gen)
    rho_ret = sum(em_ret) / len(em_ret)
    if rho_gen + rho_ret == 0:
        return 0.0
    return (rho_gen - rho_ret) / (rho_gen + rho_ret)

Common pitfalls

• Length discrepancy between generated and retrieved contexts can bias merging; the authors enforce a strict length constraint (<3% difference) to isolate semantic/conflict effects.
• Parametric knowledge may still influence answers despite traceability filtering; the paper notes complete elimination is challenging but shows negligible impact on conclusions.
• The metric assumes the model selects one context; instances where the model answers correctly from neither context are considered negligible but can skew ρ values if frequent.

Evidence (verbatim from paper)

To facilitate a simple and efficient experiment, we define a synthesized metric as follows: DiffGR = (ρ_gen - ρ_ret) / (ρ_gen + ρ_ret). The metric DiffGR, ranging from [-1,1], quantifies the extent of LLMs’ tendency to rely on generated contexts over retrieved contexts.

Citation

@misc{tan2024blinded,
  title={Blinded by Generated Contexts: How Language Models Merge Generated and Retrieved Contexts When Knowledge Conflicts?},
  author={Hexiang Tan et al. (2024)},
  year={2024},
  note={arXiv:2401.11911}
}

• arXiv: 2401.11911