Measuring and Using a Model's Own Recognition Advantage at Each Hop

Language models are often much better at recognizing a correct answer among candidates than at producing it from scratch. This is the discriminative-generative gap. In multi-hop question answering, where a system must chain facts across several retrieved documents ("who directed this film, and where was that person born?"), that asymmetry looks like a free verifier. The model could generate a hop, then check it in recognition mode before moving on. The catch is that the checker reads the same cluttered context that misled the generator. Irrelevant retrieved passages are known to pull models toward confident wrong answers, so the recognition advantage may vanish exactly where it is needed.

This note proposes to measure that advantage hop by hop: how it changes as distractor passages are added, and whether it survives when memorized world knowledge is taken away. It then proposes a self-verifier built to keep the advantage alive. The verifier is the generator itself, scoring each intermediate claim against only the single sentence the claim cites, never the full accumulated context. It intervenes only in conditions where the measurement says recognition actually beats generation. If the recognition advantage is real, cheap, and robust to noise when the verifier's view is narrowed, then small models can guard their own reasoning chains without a frontier judge. If it collapses under noise, current "use the model to check itself" designs rest on a false premise, and that is worth knowing.

Core Claim

The claim has two parts. The first is a measurement prediction. For a fixed model, define the hop-level gap as

$$\Delta(n, c, v) = D(n, c, v) - G(n, c),$$

where $G$ is the probability that the model's freely generated bridge entity at a hop is correct, and $D$ is the probability that the same model, in scoring mode, ranks the correct bridge claim highest among a candidate set. Both are evaluated with $n$ injected distractor passages, on factual ($c=0$) or counterfactual ($c=1$) chains, with the verifier given context view $v$.

We expect three things. When the verifier sees the full accumulated context, $\Delta$ is positive at low noise and shrinks toward zero or below as $n$ grows. When the verifier sees only the cited sentence, $\Delta$ stays positive across $n$. On counterfactual chains, full-context verification degrades further, because the model's memory of real-world facts votes against the passage.

The second part is a method claim. An in-loop self-verifier that is local in its view, trained on counterfactual chains, and gated by the measured gap will improve both final accuracy and chain validity on noisy multi-hop benchmarks. It will do so at lower cost than an external judge, and it will beat the simpler fixes listed below at matched compute.

What is new relative to the closest work is the combination of four things: the verifier is the generator itself, its view is restricted to cited evidence, it is grounded by counterfactual training, and it comes with an explicit measurement of when self-verification is worth doing. GRAT already uses process rewards inside a tree search over multi-hop reasoning paths, but it is trained and evaluated on factual Wikipedia benchmarks and does not ask whether its reward reads context or memory. DiVA pairs a generative search agent with a separate discriminative scorer, but only to judge completed claims after the fact. AgentEval shows that judging a step against its immediate upstream context beats judging it against the whole trace, but uses GPT-4o offline for debugging, not during inference.

Research Program

Motivating evidence. The proposal rests on three lines of prior evidence. WorldCuisines and Commonsense Reasoning in Arab Culture document large drops when models must produce rather than select answers; this supplies the recognition advantage we hope to use. Improving Zero-shot Reader by Reducing Distractions shows readers force answers out of irrelevant documents rather than abstaining; this is the reason the advantage might not survive noise. CofCA rewrites multi-hop passages with counterfactual facts and finds GPT-4 produces only 36.3% fully valid reasoning chains, with exact match dropping from about 70% on factual data to about 53% on 2-hop counterfactual data. This tells us that standard benchmarks cannot distinguish a verifier that reads the passage from one that remembers the answer.

Gap measurement. We first build a hop-level probe from MuSiQue, which composes multi-hop questions from single-hop pieces and so provides gold intermediate answers (MuSiQue). For each hop we inject $n \in \{0, 5, 10, 25\}$ distractor passages; the 25-chunk ceiling follows Ext2Gen. We build candidate sets from the gold bridge plus entities extracted from the distractors and from the model's own sampled generations, so the "hard negatives" are exactly the errors noise produces.

We measure $G$ by greedy generation. We measure $D$ by the model's length-normalized likelihood that a claim of the form "the bridge is X, as stated in sentence S" is supported, under two views: the full accumulated context, and the cited sentence alone. We repeat the measurement on counterfactually rewritten versions, built by entity swaps on MuSiQue training chains in the style of CofCA, and on CofCA itself, which is held out for testing only.

Method. We run an iterative retrieve-and-reason loop in the style of IRCoT. At each hop the model samples a small number of candidate steps, each required to cite one retrieved sentence. The same model then scores each candidate against its cited sentence alone, and we keep the highest-scoring candidate. When every candidate scores below a threshold, the system erases the hop and re-retrieves, borrowing the retry-from-last-valid-state idea of ERL. ERL detects faulty steps with TF-IDF and F1 overlap to gold annotations; we replace that with a semantic check that needs no gold answer at inference.

The verifier is lightly tuned from final-answer labels alone using the asymmetric "weakest link" aggregation of Kim's discriminative PRM: the minimum over step scores for correct chains, the product for incorrect ones. Kim tested this only on math; we transfer it to retrieval chains and train it on counterfactualized data so that memorized facts cannot satisfy the objective.

Gating. Using the measured $\Delta$ curves, the loop calls the verifier only in regimes where $\Delta$ was positive on a development split. Otherwise it falls back to plain generation. This makes the measurement operational rather than descriptive.

The rationale question. We address the unresolved tension between Con-J, which finds that written rationales protect judges from length and format shortcuts, and Kim, who finds scalar discriminative verifiers faster and harder to game. We run a rationale-first variant of the self-verifier, with the rationale conditioned on the same cited-sentence view. We ask whether the rationale helps under noise or just reintroduces distraction and latency.

Baselines

Naive reference. The do-nothing reference is single-shot retrieval followed by direct answering, and plain IRCoT with no verification. Both use the same base model (Llama3-8B and Llama-3.1-8B-Instruct), the same retriever and corpus, and the same distractor injection. The file gives no isolated absolute IRCoT number on these benchmarks; IRCoT itself reports retrieval recall gains of 11 to 21 points and QA F1 gains of up to 15 points over one-step retrieval. We will rerun it rather than quote it.

Strongest search-based competitor. This is GRAT: tree search over reasoning paths guided by process rewards. It reports 75.2% accuracy on 2WikiMultiHopQA with Llama3-70B. With Llama3-8B fine-tuned on 11,000 GRAT trajectories, it reports a rise from 60.6% to 66.8% on 2WikiMultiHopQA. The 8B figure is the fair target. We will rerun GRAT at 8B with the same number of verifier and model calls per question as our method, the same retriever, and the same hyperparameter search budget. Beating it means exceeding its accuracy on 2WikiMultiHopQA at equal calls, and, more importantly, exceeding it by a clear margin on CofCA chain validity and under $n=25$ distractors, where we predict its reward degrades.

Strongest external judge. This is a prompted GPT-4o step judge given the same local view; AgentEval's local-context design is the template. On MuSiQue, DiVA reports 79.87% Precision@1 against 57.14% for a naive generative verifier. We will report DiVA's post-hoc ranking on its own FGVeriBench as a reference, without claiming direct comparability. For the external judge, success means matching its step accuracy within a few points at a small fraction of its cost.

Noise-robust single-pass extractor. This is Ext2Gen, which reports an average score of 0.557 against 0.473 for the default Llama-3.1-8B-Instruct across noisy settings. We run it at 25 chunks on the same questions.

Ablation of the key ingredient. The ablation isolates the central idea. It is the identical self-verifier given the full accumulated context instead of the cited sentence, holding training, sampling budget and gating fixed. If this ablation matches the local verifier, the core claim is wrong.

Why not just…?

Why not just use self-consistency, sampling several chains and taking the majority answer? This is the most serious threat, because it costs nothing to build. Majority voting removes random errors. Distractor-induced errors, however, are systematic: the zero-shot reader study shows models confidently forcing answers from the same irrelevant passages. Systematic errors should concentrate votes on the wrong bridge. We will run self-consistency at exactly matched sample counts. If it matches our method on CofCA under heavy noise, the verifier adds nothing, and we will report that.

Why not just filter documents before each hop with a relevance reranker, as HANRAG does? HANRAG uses a relevance discriminator to discard irrelevant chunks before generation and reports nearly 7% exact match gains on MuSiQue over Adaptive-RAG. Document filtering cannot catch the case where the right passage is present but the model extracts the wrong entity from it. Ext2Gen also found that pre-generation compressors drop bridge entities. Still, a per-hop reranker could capture most of the gain. We include a per-hop reranker arm at matched cost, and we break down errors by type: wrong-document errors versus wrong-extraction errors.

Why not just prompt GPT-4o as the in-loop judge? It may well be more accurate per step. But tree search or best-of-many sampling multiplies judge calls per question, and Con-J reports a tuned 7B judge beating GPT-4o on Math (85.0% vs 74.8%), so frontier judges are not automatically better. We report accuracy against dollar and latency cost. A GPT-4o judge that wins on accuracy at many times the cost is a legitimate outcome, not a refutation.

Why not just train an outcome-supervised verifier on full context and skip the measurement? Kim's method already does this for math. Our full-context ablation is exactly this system transferred to retrieval. The counterfactual split tells us whether it learns to read passages or to recall facts.

Proposed Experiments

The benchmarks are MuSiQue-Ans and MuSiQue-Full (the latter for abstention), 2WikiMultiHopQA, HotpotQA in its distractor setting, Bamboogle, and CofCA as a held-out counterfactual test. Each is run at four distractor levels. The metrics are answer exact match and F1; supporting-fact F1 where gold annotations exist; full chain validity in CofCA's step-wise sense; verifier step accuracy against MuSiQue's gold intermediate answers; and model calls and wall-clock time per question.

The first decisive check is the gap curve. Plot $\Delta$ against $n$ for both verifier views, on factual and counterfactual data. The proposal stands only if the cited-sentence view keeps $\Delta$ clearly positive at $n=25$ on counterfactual chains while the full-context view does not.

The second check is the main comparison on CofCA and noisy MuSiQue against GRAT-8B, self-consistency and the per-hop reranker at matched calls. Chain validity is the primary metric, since outcome accuracy can be won by memorization. We look for a gap clearly larger than run-to-run variance across three seeds.

The third check is the memory test. The verifier's accuracy should drop far less from factual to counterfactual chains than an off-the-shelf judge's does. That is the signature of a verifier that reads rather than remembers.

The fourth check tests whether gating matters. Turning the verifier on everywhere should hurt in the regimes where the measured $\Delta$ was negative. That would show the measurement predicts where self-checking pays.

The rationale variant is reported on the same axes to resolve the Con-J versus Kim question for this setting.