Self-Targeted Distractors: When a Model's Own Guesses Steer Multi-Hop Retrieval Wrong

Many retrieval-augmented language models answer multi-step questions by looping. The model writes a partial answer, uses that text as the next search query, reads what comes back, and repeats. This works because the model's guesses often name the missing "bridge" entity, the intermediate fact a question depends on but never states. It also creates a specific risk. When the guess is wrong, the search is aimed at text that agrees with the model's mistaken belief. If plausible text like that exists in the corpus, the model finds it, reads it as confirmation, and builds a fluent, confident, wrong chain of reasoning.

We propose to measure this confirmation effect directly by building self-targeted distractors. These are misleading passages whose false content is chosen to match what the target model already wrongly believes when answering without retrieval. We then test a minimal defense. The model's guesses may shape search queries, but they may enter the reasoning chain as premises only after a retrieved passage independently supports them under the question's exact constraints. The main deliverables are a diagnostic showing which retrieval loops are vulnerable and why, and a verifier trained on these distractors. A further constraint is that the defense must not "win" by retreating into vague, non-committal answers.

Core Claim

Prior work treats plausible distractors as a property of the text. We claim the most damaging distractors are a property of the pair formed by the text and the model reading it. The more probability the model assigns, without retrieval, to the false value a passage asserts, the more that passage hurts. We further claim the harm is largest in loops that feed the model's own generation back into retrieval.

The file contains a direct tension on this point. ITER-RETGEN uses the model's intermediate output, partial hallucinations included, as an expanded query, and reports up to 8.6% absolute gains on multi-hop benchmarks. In the other direction, Seemingly Plausible Distractors in Multi-Hop Reasoning shows that synthetic paragraphs supporting a coherent false reasoning path cut LLM performance by up to 45%. It also finds that chained false paragraphs (the "relatedness trap") hurt more than isolated ones, and that chain-of-thought prompting often makes things worse. Neither paper asks whether the false path happened to agree with the reader's own prior.

Our hypothesis reconciles the two findings. A hallucinated bridge is a helpful search cue when the corpus contains no text confirming it. It becomes a trap when such text exists. That is exactly the situation self-targeted distractors create, and it is the situation that low-quality or adversarial web content creates in deployment.

Research Program

Constructing self-targeted distractors. The construction follows the three-stage recipe of Seemingly Plausible Distractors. First, a multi-hop question is decomposed into sub-questions; MuSiQue supplies gold decompositions. Second, a modifier in one sub-question is mutated, for example a year or a qualifying noun, so that the distractor describes a near-neighbor entity rather than the true one. Third, a Wikipedia-style paragraph is written for that near-neighbor. Because of the mutation, the gold answer stays unambiguous for an attentive reader.

The new ingredient is how the false bridge value is chosen. We run the target model closed-book (without retrieval) on the mutated sub-question and rank candidate wrong values of the correct entity type by the model's own likelihood. This borrows the perplexity-guided selection of Are the Hidden States Hiding Something?, which isolates false statements a model finds most plausible. It also follows the observation in Lelkes et al. that closed-book answers make highly plausible distractors (61.8% of their questions had at least one).

Each self-targeted distractor gets a matched control. The control is produced by the same mutation and the same paragraph generator, but it plants a type-matched value the model assigns low probability. The two passages differ only in whether they agree with the reader's prior. This lets us estimate a dose-response curve: accuracy drop as a function of the closed-book log-probability of the planted value.

Decomposing the failure. For each pipeline we split the harm into two parts. Retrieval pull is the rate at which the distractor enters the retrieved context. Reading adoption is the rate at which the final chain uses the planted value, given that the distractor was retrieved. Our hypothesis predicts that prior-alignment raises retrieval pull mainly in generation-as-query loops such as ITER-RETGEN. It predicts a raise in reading adoption in all pipelines. The decomposition matters because the two parts call for different fixes.

The defense: guesses may search, only evidence may assert. We modify an interleaved loop in the style of IRCoT to keep two channels. A scratch channel holds the model's guesses and is used only to form search queries. The reasoning channel may contain a bridge fact only if a retrieved passage passes a step verifier. The verifier is a small model, and it checks whether the passage asserts the relation for the entity the sub-question actually describes, all modifiers included. In other words, it checks the attentive reading that the distractor mutation is designed to defeat.

The verifier is trained on self-targeted distractors generated against a pool of open models, with original true paragraphs as positives. This is a retrieval-grounded, open-text analogue of the adversarial step-error generation in Adversarial Training for Process Reward Models, which works only on math and science without retrieval. Unlike DPRM, it needs no pre-existing knowledge graph to check steps against. Its explicit apply/ignore judgment resembles Reason-KE. The difference is that Reason-KE receives facts and distractors upfront in its context, while our verifier operates inside open-corpus iterative retrieval.

The vacuity constraint. Any hallucination defense can trade wrong answers for empty ones. A system that answers "I could not determine this" whenever passages conflict reproduces the generic response problem from dialogue. That problem is the tendency of likelihood-trained models to fall back on safe, low-information outputs. Conversing by Reading traces it to the fact that such outputs are acceptable across many contexts, and its CMR model counters it by grounding responses in documents.

We therefore score every system on answerable questions jointly for accuracy and answer rate. We count abstention or hedging on answerable items as failure, not safety. We also run a secondary check on the Wizard-of-Wikipedia dialogue data used by AKGCM. The check asks whether evidence gating, when transplanted into a grounded chatbot, lowers specificity, measured by distinct-n and grounding F1 as in CMR.

Baselines

The naive reference point is the same model with one-shot dense retrieval and no defense, run on clean and distractor-injected corpora. It indicates how often distractors are even reached. QASC found that single-step retrieval recovers both needed facts only 2.9% of the time when the bridge concept is unstated, so this baseline may look "robust" only because it retrieves little. We will report its retrieval pull separately to expose this.

The iterative loops are IRCoT and ITER-RETGEN. On clean benchmarks, IRCoT reports 11–21 point retrieval recall gains and up to 15 QA F1 points, and ITER-RETGEN reports up to 8.6% absolute gain. These loops are the systems whose vulnerability we predict, and they are the base our defense modifies. To keep comparisons fair, each runs with the same reader model, the same retriever and index, the same cap on retrieval rounds (three, since ITER-RETGEN plateaus within two to three), and the same per-question token budget.

The strongest existing defense is a Reason-KE-style relevance filter. A model is fine-tuned to label each retrieved passage apply or ignore before answering. Following Reason-KE, its training data has 2–4 injected generic plausible distractors per question. It is then placed inside IRCoT. It is the strongest competitor because it already reports large gains in distractor resilience: 90.2% multi-hop accuracy on MQuAKE-CF for Qwen2.5-7B, with only a 6.3% drop under heavy distraction. That figure comes from a closed setting where facts are given upfront. For open-corpus HotpotQA or MuSiQue with self-targeted distractors, we found no reported result on Lacuna.

For fairness, we train it on the same base model, the same number of training questions, and the same hyperparameter-search budget as our verifier. Beating it means higher answer accuracy on the self-targeted test sets at matched answer rate, with no loss on clean data. Gains that appear only because our system abstains more do not count.

We also include uncertainty-triggered retrieval in the style of DRAGIN (Su et al.). It fires retrieval when token-level uncertainty is high. We predict it misses exactly the self-aligned case, because the model is confident there.

The key ablation isolates the novel ingredient. Our verifier is trained three ways that differ only in distractor source: self-targeted distractors, matched low-prior controls, or none. A separate ablation removes the two-channel separation, letting guesses enter the reasoning chain directly while the verifier is kept.

Why not just…?

Why not just train on the existing generic plausible distractors? This is the most serious threat. If verifiers trained on the matched low-prior controls defend as well as verifiers trained on self-targeted ones, the "self" in our claim adds nothing for defense, and the contribution shrinks to a diagnostic. The training-source ablation settles this directly, since both variants share every component except the planted value. We will report the outcome either way.

Why not just stop feeding generations into the query, that is, use IRCoT instead of ITER-RETGEN? If query contamination drives most of the harm, simply keeping guesses out of queries may suffice. However, IRCoT's reasoning sentences also contain the model's guesses and also become queries, so we expect only a partial reduction. The retrieval-pull versus reading-adoption decomposition will show whether removing guesses from queries addresses the larger part of the damage. If reading adoption dominates, query hygiene alone cannot help.

Why not just ask the model, or probe its hidden states, whether a step is true? Hidden-state probes collapse to 50–60% accuracy on realistic, low-perplexity hallucinations (Are the Hidden States Hiding Something?). Prior-aligned distractors are the low-perplexity case by construction. Direct self-queries invite the model to confirm its own errors (Agrawal et al.). Their indirect, multi-sample cross-examination is stronger, so we include it as an extra baseline at matched compute. We expect it to struggle here because the error is stable rather than variable: the model consistently believes the planted value.

Why not just use a bigger model? Scale does not remove distractor susceptibility. GPT-4 still lost 14 F1 points with four adversarial paragraphs in Seemingly Plausible Distractors. We will test whether the dose-response slope against the model's own prior shrinks with size across the 7–70B models we evaluate.

Proposed Experiments

The test beds are HotpotQA and MuSiQue. On HotpotQA, the original distractor-setting baseline answer F1 is 58.28 against human 91.40. MuSiQue's connected 2–4 hop construction resists shortcut reasoning; human performance there is 78.0 answer F1. 2WikiMultihopQA is a third test bed. For each benchmark we inject one self-targeted chain or one matched-control chain per question into the Wikipedia paragraph pool. Distractors are generated separately for each target reader, for example a Llama and a Qwen model at 7–8B and 70B. Cross-model transfer, meaning a distractor targeted at model A and read by model B, gives a further control on whether the effect is truly "self."

Human annotators check a sample of injected passages to confirm that the gold answer stays unambiguous. This guards against the accidentally-correct distractors noted by Lelkes et al.. The metrics are answer exact match and F1, answer rate, retrieval pull, reading adoption, and the rate of planted-value adoption in reasoning chains.

Three checks decide the outcome. First, on undefended pipelines, accuracy should fall more under self-targeted distractors than under matched controls, and the gap should be largest for ITER-RETGEN. A flat dose-response curve would falsify the core claim. Second, our gated verifier should beat the Reason-KE-style filter on self-targeted test sets at matched answer rate, and the advantage should vanish when our verifier is trained on matched controls instead. Third, on clean benchmarks and in the Wizard-of-Wikipedia transfer, the gated system should lose at most a small margin in accuracy and response specificity. A defense that turns hallucination into vacuity has failed by our own standard.