Forgetting the Path, Not Just the Fact: Model-Verified Chains for Multi-Hop Unlearning

When a language model is made to "forget" a fact, such as who runs a company, it usually stops answering the direct question. It can still answer an indirect one. Asked "where was Tesla's CEO born?", the model silently recovers the forgotten link and uses it as a stepping stone to the answer. Unlearning here means removing the influence of specific data from a trained model's weights. Breaking Chains shows this leak clearly. On Llama-3.1, a standard unlearning method (Negative Preference Optimization with a retain loss, NPO+RT) lowers multi-hop probing accuracy only from 98.1% to 90.0%. The paper's fix is an inference-time gate that rejects suspicious queries, and its authors explicitly call for methods that remove the multi-hop pathways from the parameters themselves.

We propose a way to do that. The method trains the model to forget on multi-hop questions that route through the target fact. It keeps only the questions the model's own pre-unlearning answers confirm it actually reasons through. Training is posed as a constrained problem: every neighboring fact on those chains, such as "Musk was born in Pretoria", must stay intact. The bet is that multi-hop leakage after weight-level unlearning is mainly a data-selection problem rather than an optimizer problem. Synthetic chains that are hallucinated, or that the model never knew, waste the forgetting budget and damage nearby knowledge. Chains checked against the model's own beliefs, paired with explicit protection of sibling facts, should break the composition without collapsing the model.

Core Claim

Augmenting the forget set with multi-hop questions is the obvious move. Done naively, it either fails or does damage, for two reasons that can be measured separately.

The first is unknown chains. A generator such as GPT-4o produces chains whose intermediate links the model does not hold, so pushing down their likelihood teaches nothing about the target edge.

The second is collateral damage. The augmented questions also contain the other hop. Without a constraint, optimization pushes the model to forget "Musk was born in Pretoria" along with "Musk is Tesla's CEO". This is the collateral forgetting that HAMU formalizes through the inner product of the forget and retain gradients.

Our claim has two parts. First, filtering chains by model-grounded consistency should recover most of the multi-hop forgetting that MUNCH achieves at inference time, without its runtime cost. A chain counts as consistent when the pre-unlearning model answers each hop correctly and its composed answer matches its direct multi-hop answer. Second, per-edge retain constraints should keep single-hop recall of sibling facts within a fixed budget, where unfiltered augmentation does not.

Research Program

The pipeline has three stages.

Generating candidate chains. For each target fact, stored as a subject-relation-object triple, we expand a local knowledge graph by querying the pre-unlearning model, following the probing procedure of The Unlearning Mirage. We then compose 2–4 hop questions whose reasoning path passes through the target edge. We also compose sibling questions that use the same bridge entities but avoid that edge.

Filtering chains. Each candidate is admitted only if two independent routes agree, adapting the agreement filters used in synthetic-data work to the model's own beliefs. In TableLLM, language reasoning and code execution must reach the same value. In Promptagator, a synthetic query must retrieve its source document. Our two routes are the hop-by-hop decomposed answer and the direct end-to-end answer of the pre-unlearning model. A second check runs a held-out model on the same question as a cross-model test, in the style of Ensemble-Instruct. This check separates "the model knows this chain" from "this chain is true in the world". We keep both labels, because forgetting should target what the model knows, and retain constraints should protect only true facts.

Constrained optimization. Let $\theta$ be the model weights. Let $\mathcal{L}_F(\theta)$ be the NPO loss on the filtered forget chains, using the objective from Negative Preference Optimization. Let $\mathcal{L}_{e}(\theta)$ be the cross-entropy on single-hop and sibling-chain questions for each protected edge $e$. Let $\epsilon_e$ be the allowed increase over the original model's loss on that edge, written $\mathcal{L}_{e}(\theta_0)$. We solve

$$\min_\theta \; \mathcal{L}_F(\theta) \quad \text{s.t.} \quad \mathcal{L}_{e}(\theta) - \mathcal{L}_{e}(\theta_0) \le \epsilon_e \;\; \text{for every protected edge } e,$$

using primal-dual updates with one dual variable $\lambda_e$ per edge group. The dual variables act as dynamic regulators, as in Jiang et al.: each $\lambda_e$ grows when edge $e$ starts to erode and relaxes once the edge is safe. Per-edge duals matter because, as HAMU argues, collateral damage concentrates on edges close to the forget target rather than spreading evenly across the retain set.

An LLM judge is used only for measurement, not training. A GPT-4o judge reads a free-form multi-hop answer and labels each step as leaking the target or not, with step-level labels in the spirit of AgentEval. We calibrate the judge against 300 human-labeled traces before trusting it, because Leak-Resistant Unlearning warns that LLM-based checking is vulnerable to bias and shortcuts.

Baselines

The do-nothing reference is the original model. It establishes ceiling leakage and utility, for example 98.1% multi-hop accuracy on the Breaking Chains Llama-3.1 setup.

The standard weight-level reference is NPO+RT trained only on single-hop forget facts. On this method, Breaking Chains reports multi-hop accuracy of only 90.0%.

The strongest parametric competitor is NPO+RT with unfiltered multi-hop augmentation: the same generator, the same number of forget examples, and the same retain set as our method. It isolates whether filtering and constraints add anything beyond "more multi-hop data". We also run HAMU's rectified update on the same data as a constrained-optimization competitor that has no notion of edges. We run RMU from WMDP, which preserves utility well; WMDP reports MMLU moving from 58.1% to 57.1% on Zephyr-7B, but Leak-Resistant reports that RMU fails to remove target knowledge.

The strongest method overall is MUNCH, the inference-time gate from Breaking Chains. It reduces multi-hop accuracy to 6.8% while keeping MMLU, at the cost of running a teacher model and a pre-unlearning model at inference time. We also report PoRT where its code allows.

All parametric methods share the same base models, forget and retain pools, number of optimizer steps, and a matched hyperparameter search of 20 configurations each. Each method is selected on a validation split whose probes come from a different pipeline than its training chains.

Two ablations isolate the key ingredients. One removes the consistency filter but keeps the constraints. The other keeps the filter but replaces per-edge duals with a single fixed-weight retain loss.

Beating the baselines means the following. We must reach held-out multi-hop leakage within 10 points of MUNCH's 6.8% with no inference-time overhead. We must keep sibling-edge single-hop recall within $\epsilon$ of the original model, where unfiltered augmentation does not. We must also show lower recovery under Leak-Resistant's attacks than MUNCH's underlying unlearned weights, which still hold the knowledge.

Why not just…?

Why not just augment the forget set with GPT-4o multi-hop questions and run NPO? This is the real threat, and the proposal is built to test it directly. If unfiltered augmentation matches our method on both held-out leakage and sibling-edge recall, the filtering and constraint ingredients are unnecessary, and we will report that. Our prediction rests on the two failure modes in the Core Claim, unknown chains and collateral damage on shared hops. We will measure both, reporting the fraction of generated chains that the pre-unlearning model fails on and the change in recall per sibling edge.

Why not just use MUNCH or PoRT at inference time? Both work against direct leakage: MUNCH reaches 6.8%, and PoRT holds WMDP accuracy near the 24.9% chance level. However, they leave the knowledge in the weights. Intrinsic Evaluation of Unlearning Using Parametric Knowledge Traces shows that internal concept vectors survive fine-tuning-based unlearning, with top-token Jaccard similarity above 0.98, and that jailbreaks can amplify them. Leak-Resistant shows that sampling and quantization recover supposedly deleted facts. Our recovery-attack experiments compare the weights directly, which a gate cannot defend against.

Why not just train longer or raise the forget weight? Breaking Chains reports that more data or epochs causes catastrophic forgetting without fixing multi-hop leakage. NPO documents the collapse of gradient ascent into repetitive gibberish. We include an epoch sweep of the single-hop NPO+RT baseline to confirm this on our splits.

Why not let an LLM judge choose the forget chains instead of model-grounded consistency? A judge can say whether a chain is true in the world. It cannot say whether this particular model reasons through that chain, and that is the property forgetting needs. As a fourth arm, we run judge-filtered augmentation with the same data budget.

Proposed Experiments

Models and benchmarks. We use Llama-3.1-8B, following Breaking Chains, and a second open model. Benchmarks come from three sources. The first is the MQuAKE-derived Breaking Chains setup built on MQuAKE. The second is the Leak-Resistant suite, which has 3,000 MQuAKE questions across logical structures such as Modus Tollens and Disjunctive Syllogism, plus three recovery attacks. The third is RWKU probes from the Mirage framework, where multi-hop accuracy reportedly stays at 20–30% after existing unlearning methods.

Utility and forget quality. We measure utility on MMLU and BBH; Leak-Resistant reports that PALU halves BBH. We measure TOFU forget quality and model utility, using the p-value criterion of TOFU.

Metrics. Metrics are held-out multi-hop leakage scored by the calibrated judge, exact match as a cross-check, recall per sibling edge, recovery rate under each attack, and utility.

Decisive checks. The first check is the filter ablation against unfiltered augmentation at matched data size. If the gap is under 3 points of leakage and recall, the central claim fails. The second is per-structure analysis on Leak-Resistant. If our method closes the Modus Tollens and Disjunctive Syllogism gaps that Leak-Resistant reports, for example Task Vector falling from 63.5% on Hypothetical Syllogism chains to 27.6% on Modus Tollens, that shows path-level forgetting rather than template memorization. The third is a held-out-generator test. Training chains come from our pipeline and test probes come only from the Mirage and Leak-Resistant pipelines, so gains cannot come from shared phrasing.