Do Step-Level Judges Read the Evidence? Interventional Audits for Multi-Hop Explanations

When a language model answers a question by chaining several retrieved documents, as in "in what city was the director of Inception born?", the usual way to check it is to ask a second, stronger model to grade each intermediate step. This practice is called LLM-as-a-judge. A grade of "this step is supported" is only useful if the judge reached it by reading the retrieved evidence. A judge that already knows Christopher Nolan was born in London can approve the step whether or not the trace ever retrieved that fact. We propose to test this directly. We edit the evidence under a step, by deleting it or by swapping in a counterfactual version, and measure whether the judge's verdict changes when it should and stays put when it should not. We then build a judging protocol that only credits a step when its grade survives those edits.

The stakes are highest for "why" and "how" questions. These have no answer type, such as a name or a date, that a string-matching metric could check, so a model judge is the only scalable evaluator. They also require chaining implicit cause-and-effect links across passages. The thing that most needs a judge is therefore also where judges are least audited.

Core Claim

Current step-level judges for multi-hop reasoning have never been tested for whether their verdicts depend on the evidence a step cites. We expect two failures in practice. First, judges will credit steps from their own memory. Second, they will penalize evidence-faithful steps when the evidence contradicts what they memorized.

The proposal makes two contributions. The first is an audit that measures both failures through controlled edits to the retrieved evidence. The second is counterfactual-contrast judging. In this protocol a step counts as supported only if two conditions hold. Its grade must drop when its cited parent evidence is removed. Its grade must also track the edit when the bridging fact in that evidence is changed.

The idea borrows its logic from two sources. One is the interventional test of A Causal Lens for Evaluating Faithfulness Metrics, which edits model knowledge to see whether faithfulness metrics tell faithful from unfaithful explanations. The other is the ancestor-crediting scheme of PRISM-Physics. We apply both to the judge itself, on open-ended text, inside retrieval pipelines.

What Prior Work Establishes and Leaves Open

Three findings motivate the work.

The first is about multi-hop reasoners. CofCA rewrites Wikipedia passages with fictional entities and shows that reasoners fall back on pre-training memory. GPT-4's exact match drops from about 70% on factual passages to about 53% on 2-hop counterfactual ones, and only 36.3% of its reasoning chains are correct. Nothing suggests that a judge built from the same kind of model is immune.

The second is about judges. The Silent Judge shows that judges shift pairwise verdicts by up to +30% for a "2025" versus "1950" tag on ELI5. It also shows a 0% rate of acknowledging that tag, so a judge's written rationale cannot be trusted to reveal what drove its verdict. That study was single-turn. Multi-hop retrieval traces naturally carry the same date and source cues in document metadata.

The third is about faithfulness metrics. The Causal Lens paper finds that every faithfulness metric it tested breaks down on multi-hop StrategyQA. It diagnoses the failure but proposes no repair.

Two papers supply the methods we compete with.

AgentEval represents agent runs as a dependency graph, a DAG whose nodes are steps and whose edges say which outputs feed which steps. It grades each step with GPT-4o against its upstream context and traces errors back to their source. It reports failure-detection recall of 0.89 against 0.41 for end-to-end evaluation, Cohen's κ of 0.84 with experts, and 72% root-cause accuracy. It never edits upstream content to check that the judge's step grades actually depend on it.

RFEval applies edits to the reasoner rather than the judge. It finds 49.7% unfaithful outputs, but it relies on an unaudited o3 judge to read those outputs.

One paper constrains the design. PRISM-Physics rejects LLM judges altogether in favor of symbolic checks. Its DAG scoring reaches Kendall's τ_b of 0.346 with human graders, against 0.294 for an LLM judge. Symbolic checks are unavailable for textual causal chains. So we keep the judge but subject it to tests that play the role of PRISM's equivalence check.

Grounding Multi-Hop Reasoning in Structural Causal Models via GRPO shares our graph view of evidence and inference. It trains a reasoner against gold labels and explicitly does not perform interventions or use a judge.

Research Program

Trace construction

We build traces with dependency graphs we can trust.

MuSiQue (♫ MuSiQue) composes each question from single-hop sub-questions and filters out shortcuts. This gives a gold dependency graph for 2–4 hop questions.

For the causal regime, we draw on CausalQA's Webis-CausalQA-22. It extracts causal questions from ten sources, including HotpotQA, by lexical rules. We take the HotpotQA-derived causal questions, whose supporting-sentence annotations from HotpotQA give partial evidence graphs. We supplement these with roughly 300 ELI5 causal explanations, which annotators segment into cause-to-effect chains with cited passages.

CofCA's counterfactual passages are placed into a retrieval pool with distractors. This turns CofCA's closed reading-comprehension setting into a retrieval one.

We then run three reasoners over these corpora:

  • a one-shot dense retriever with a reader;
  • IRCoT, which interleaves chain-of-thought sentences with retrieval;
  • a tree-search variant in the style of GRAT, if its implementation can be reproduced.

The outputs are traces in which each step cites the passages it uses.

Interventions with known ground truth

For each step we apply four edits. Each comes with a label that holds by construction.

  1. Parent removal. Deleting the step's cited parent passage should make the step unsupported in context, even though it remains true in the world.
  2. Counterfactual substitution. Changing the bridging entity or causal link in the parent, CofCA-style, should make the original step unsupported and make the edited claim supported.
  3. Non-parent edit. Changing a passage the step does not depend on should leave the verdict unchanged.
  4. Cue injection. Following the Silent Judge design, we alter only the date or source metadata on retrieved documents.

These edits yield four audit quantities:

  • Intervention sensitivity: the fraction of edits 1 and 2 that flip the verdict correctly.
  • Collateral flip rate: the fraction of edit 3 that flips the verdict.
  • Verdict shift rate under edit 4.
  • Cue acknowledgment rate under edit 4.

Counterfactual-contrast judging

Let $J(s \mid C)$ be the judge's probability that step $s$ is supported in context $C$. Let $C^{-}_s$ be $C$ with the step's parents removed, and $C^{\ast}_s$ be $C$ with the substituted parents. The contrast score is

$$S(s) = \min\big(J(s \mid C) - J(s \mid C^{-}_s),\; J(s \mid C) - J(s \mid C^{\ast}_s)\big).$$

A step is credited only if $S(s)$ exceeds a threshold tuned on a held-out split. Credit then propagates along the dependency graph in the manner of PRISM's ancestor closure.

The protocol costs about three judge calls per step. Removal and substitution are generated automatically. The substitutions use an LLM rewriter followed by an entailment filter, with human spot checks on a 5% sample.

The scope is limited in two ways. First, the method needs traces whose steps cite passages. Second, for open-ended causal steps it depends on how cleanly a causal link can be swapped. The ELI5 human chains test that directly.

Baselines

End-to-end outcome judgment. This is the do-nothing reference: a single judge call per trace that grades only the final answer, or exact match where answers are entities. The same judge model sees the same traces. Beating it means higher step-level detection of unsupported steps. AgentEval reports 0.41 recall for this style of evaluation on its agent traces. No reported result was found on Lacuna for end-to-end judging of MuSiQue or CofCA traces at the step level.

AgentEval-style step judging. This is the strongest existing competitor. We reimplement AgentEval's protocol: GPT-4o grades each step against its immediate upstream context using its failure taxonomy, and errors are traced back along the dependency graph. It is the published method closest to ours in structure. Its reported numbers are recall 0.89, κ 0.84 and 72% root-cause accuracy. All three are on agent workflows, not multi-hop question answering.

To keep the comparison fair, it receives the same judge model, the same traces and the same prompt-tuning budget. It also gets three samples per step with majority voting, so it matches our call count. Beating it requires two things on the intervention-labelled test set:

  • higher intervention sensitivity at an equal or lower collateral flip rate;
  • equal or better agreement with human step labels on the 300 annotated traces.

Ablations. Three ablations isolate the key ingredient. The first keeps only the removal contrast. The second keeps only the substitution contrast. The third replaces both contrasts with a "cite your evidence" prompt at equal call count. The third ablation separates "more compute" from "interventional checking."

Why not just…?

Why not just give the judge the gold answer and gold supporting facts? Reference-guided judging checks whether a step is true, not whether the trace grounded it. A step that is correct but unsupported in its own retrieved evidence would still pass. On CofCA-style data, gold references remove the memory conflict only for the judge, not for the system being judged.

More fundamentally, open-ended causal explanations have no gold reference to check against. CausalQA's best fine-tuned system reaches only 0.48 ROUGE-L, and the paper reports that search engines give contradictory answers to identical causal queries. We still run reference-guided judging as an oracle upper bound on the factoid subsets.

Why not just score each step with an entailment model against its cited passage? This is a real threat on MuSiQue. Its steps are often single-sentence factoid hops that entailment models handle well, and entailment ignores parametric memory by construction. The weakness is causal steps. Deep Read notes that causality is usually implicit rather than marked by "because." We expect entailment checkers to miss those links.

We include an entailment baseline at matched cost and report results separately by subset. If it matches counterfactual-contrast judging on the causal subsets as well, we will report that cheaper finding as the result.

Why not just use a stronger reasoning judge? RFEval found that scale did not buy faithfulness in reasoners. The gpt-oss-120B model was less faithful than its 20B sibling. RFEval also relied on o3 without auditing it. So we include o3-class and smaller judges in the audit itself. If the strongest judge already shows high sensitivity and low collateral flips, the contrast protocol is unnecessary for it, and the audit becomes the contribution.

Proposed Experiments

The decisive first check is the audit itself. On factual MuSiQue traces, we test whether standard step judges keep crediting steps after parent removal at rates well above zero. That would be parametric leakage, the judge-side analogue of CofCA's knowledge conflict. On counterfactual CofCA traces, we test whether they reject evidence-faithful steps that contradict memory. If neither failure exceeds a few percent, the premise is wrong and we stop.

Second, we compare protocols. Counterfactual-contrast judging is compared against all baselines on four measures:

  • intervention sensitivity and collateral flip rate;
  • step-level F1 and κ against human labels;
  • root-cause attribution accuracy along the gold dependency graph;
  • agreement with expert rankings of whole traces, measured by Kendall's τ_b so it is comparable to PRISM-Physics's reported 0.294 for an LLM judge.

Third, we test cue robustness. We inject recency and source tags into document metadata and check whether the contrast score reduces verdict shift. Silent Judge's single-turn reference point is GPT-4o's +30% on ELI5. A metadata tag that alters no content should leave both contrast terms roughly unchanged, so the score should be largely immune.

Finally, we run a downstream check. We re-rank IRCoT and one-shot retrieval systems by each judge. We then ask whether the contrast judge's system ranking flips between factual and counterfactual corpora less than the baseline judge's does. A judge that grades evidence rather than memory should rank systems the same way regardless of whether the facts are real.