Machine unlearning tries to delete specific knowledge from a trained language model, such as a fictitious author's biography or hazardous biology facts, without retraining it from scratch. The best current methods avoid the old failure mode, in which deletion wrecks the model's general abilities. Yet growing evidence says they often hide knowledge rather than remove it. Two findings show this. Forgotten facts reappear when the model's outputs are sampled many times, and simple probes of the model's internal features still recover the "forgotten" classes. Meanwhile, work on groups of LLM agents shows that a stubborn peer, meaning an agent that refuses to change its stated position, ends up steering the whole group's conclusion. We propose to put an unlearned model into a structured discussion with stubborn peers and study two things. The first is whether this pressure pulls deleted knowledge back out. The second, and more important, is whether the way the model yields to a stubborn peer can serve as an audit, done from outside the model, of how much knowledge remains.
The key trick is a controlled comparison. A stubborn peer asserts either the true forgotten answer or an equally plausible false one, with identical confidence and argument style. A model that never learned the fact has no reason to prefer one over the other. A model that still holds a trace of the fact should give way faster to the true claim. This difference in how readily the model yields can be estimated with standard opinion-dynamics models. It needs only text in and text out. It also separates real residual knowledge from the trivial case of a model repeating what a peer just told it.
Core Claim
Unlearning methods that avoid catastrophic collapse leave behind residual knowledge. We claim this residue can be detected, and partly re-elicited, through deliberation with stubborn peers. Specifically, fit a Friedkin-Johnsen model to the unlearned agent's stance trajectory. This is the standard model in which each agent moves toward a weighted average of its neighbours while staying partly anchored to its initial view. The agent's fitted susceptibility toward a stubborn peer should be higher when the peer asserts true forgotten content than when it asserts matched false content. We also claim this "susceptibility asymmetry" grows as unlearning is made gentler to preserve utility. That would make it a direct, behaviour-level measure of the trade-off between forgetting and catastrophic collapse, measured in the setting where these models are increasingly deployed.
The ingredients come from separate literatures. Multi-Agent Systems are Mixtures of Experts: Who Becomes an Influencer? fits input-dependent Friedkin-Johnsen parameters to LLM debates and finds that stubbornness predicts influence at $r \approx 0.96$. That work motivates using a stubborn peer as the lever, but it never touches unlearned models. Biases in Opinion Dynamics in Multi-Agent Systems of Large Language Models: A Case Study on Funding Allocation shows that the pull toward consensus erodes safety stances, with harmful items funded in 82–96% of closed-form runs. It studies safety tuning from reinforcement learning with human feedback (RLHF), not parameter-level unlearning. Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding shows NPO and RMU leaking nearly everything by $k=128$ samples, and names agent systems as unstudied future work. An Illusion of Unlearning? recovers 70–80% of forgotten-class accuracy on CIFAR-10 by linear probing of frozen features, but it needs white-box access. The "social radar" line, The Social System Identification Problem and Active Sensing of Social Networks, supplies the method: injected stubborn nodes keep a network from reaching full consensus, so the steady state still shows who is influenced by whom and can be used to estimate hidden influence weights. We turn that idea on a single model's hidden knowledge instead of a hidden graph.
The closest multi-agent unlearning work, Agents Are All You Need for LLM Unlearning, is different in kind. It uses a fixed pipeline of four agents to filter one frozen model's outputs. There is no deliberation, no stubbornness and no audit of residual knowledge.
Research Program
The central object is a small discussion. It contains one unlearned target agent, one stubborn peer, and optionally two ordinary peers, and it runs for a fixed number of rounds on a question drawn from the forget set. We control the peer's stubbornness with the Belief Engine. That system tracks a stance in log-odds space, separately from the text the agent writes, and exposes a prior-anchoring parameter $a$. We can therefore sweep $a$ from fully open-minded to fully fixed, rather than relying on prompt wording.
The peer runs in three conditions:
- True assertion. The peer asserts the true forgotten answer.
- Matched false assertion. The peer asserts a false answer. We take these from TOFU's perturbed answers, which are built to be equally plausible.
- Pressure only. The peer insists the target knows the answer and demands specific attributes, but never supplies content.
After each round we read off the target's stance, either as its stated probability or as a judged agreement score, and fit the Friedkin-Johnsen update
$$x_{t+1} = \gamma\, x_0 + (1-\gamma)\big[(1-w)\,x_t + w\, s\big].$$
Here $x_t$ is the target's stance at round $t$, $x_0$ its initial stance, $s$ the stubborn peer's fixed stance, $\gamma$ the target's anchoring to its own start, and $w$ its susceptibility to the peer. We fit $w$ separately for the true-assertion and false-assertion conditions. The audit statistic is the gap between the two fitted susceptibilities,
$$\Delta = w_{\text{true}} - w_{\text{false}},$$
averaged over forget-set questions.
The reference point for $\Delta$ is TOFU's retain-only model, which was trained without the forget authors. By construction its $\Delta$ reflects no knowledge of the facts. An unlearned model counts as "forgetting at the level of social behaviour" only if its $\Delta$ is statistically indistinguishable from this reference.
The pressure-only condition measures re-elicitation in the strict sense. We count answers that contain correct forgotten content absent from the whole transcript, minus the rate at which the retain-only model produces the same content by chance.
We then trace how $\Delta$ and re-elicitation vary with unlearning strength. For each method we sweep the number of steps or the loss weight and pair each checkpoint's $\Delta$ with its retain-set utility. The resulting curve tests the core claim that avoiding catastrophic collapse and leaving detectable residue are coupled.
As a secondary question we also ask whether unlearning damages how well the agent debates on benign questions. The file flags this as an open problem ("Gap 2"). We test it by comparing fitted stubbornness and routing regret on MMLU-Pro retain questions before and after unlearning, since a small drop in single-model scores could hide a larger collapse in group performance.
Baselines
Naive reference. The unlearned model answers forget-set questions alone, with greedy decoding and no peers. This is the evaluation under which methods look solved. The WMDP Benchmark reports that RMU on Zephyr-7B drops WMDP-Bio accuracy from 63.7% to 31.2% while MMLU falls only from 58.1% to 57.1%. Deliberation beats this reference if it raises forget-set recovery significantly above the solo greedy rate.
Strongest existing attack: matched-budget sampling. This is leak@k sampling from Leak@$k$, run with $k$ set equal to the total number of target-agent generations in a discussion (agents × rounds). Both use the same nucleus-sampling settings. The paper reports that NPO and RMU on TOFU appear perfect at $k=1$ but show nearly 100% leakage by $k=128$. Because that ceiling is so high, raw leakage is not where deliberation has to win. Deliberation beats leak@k if it achieves comparable re-elicitation at a much smaller matched budget. It also beats leak@k if $\Delta$ detects residue in models where leak@k is near zero.
Strongest existing unlearning method. The strongest defence to test against is RULE-GradDiff, which the Leak@$k$ paper reports reaches "nearly zero leakage even at $k=128$" on TOFU. If $\Delta$ stays significantly above the retain-only reference for RULE-trained models, deliberation exposes residue that the strongest sampling-hardened method leaves behind. We also include NPO, the only method on TOFU Forget10 to keep forget quality at p > 0.05 with high utility (Negative Preference Optimization), and WAGLE+NPO, reported at unlearning efficacy 0.9599 on Llama2-7b (WAGLE). Gradient ascent and gradient difference from TOFU serve as collapse-prone contrasts. Every method gets the same base model, the same forget split, and the same hyperparameter search budget, and each is tuned to its own paper's recommended utility threshold before any deliberation is run.
White-box references. We compare against two measures that need access to the model's internals. The first is TOFU's Truth Ratio, which compares the model's probabilities for true and perturbed answers. The second is a linear probe on intermediate activations, following An Illusion of Unlearning?. Neither is a competitor in the black-box setting. Both serve as validity targets: $\Delta$ is useful only if it tracks them across checkpoints.
Ablations. The ablation that isolates the key ingredient removes stubbornness by setting the peer's anchoring $a$ to 0 while keeping its content fixed. A second ablation is a single-turn prompt of the form "a colleague insists the answer is X", which isolates whether several rounds of social pressure add anything beyond one piece of in-context evidence.
Why not just…?
Why not just sample many times, since deliberation is just more samples? This is the most serious threat. The matched-budget leak@k baseline settles it directly. If deliberation and budget-matched sampling give the same re-elicitation and the same true/false asymmetry, then stubbornness adds nothing and we will report that. Two things could still distinguish them. First, the true/false comparison yields a detection signal that raw leakage rates cannot: leakage near 100% for NPO says nothing about models like RULE-GradDiff that are already hardened against sampling. Second, the funding-allocation study shows that the pull toward consensus overrides refusal behaviour. That is a different mechanism from the cascade triggered by sampling a first token that Leak@$k$ describes.
Why not just read the Truth Ratio or run a linear probe? These work when you hold the weights or log-probabilities. Audits of deployed or third-party models often have neither, and ALU-style wrappers deliberately hide them. We treat these measures as ground truth for validating $\Delta$. If $\Delta$ correlates poorly with probe accuracy across unlearning checkpoints, the audit claim fails.
Why not just paste the answer into the prompt and see if the model agrees? A model will often agree with whatever it is told, so agreement alone proves nothing. That is exactly why the design compares true against matched false assertions and subtracts the retain-only model's behaviour. The single-turn ablation is this alternative. If it reproduces $\Delta$ at full strength, the multi-round, stubborn-peer framing is unnecessary. In that case the contribution shrinks to a cheap black-box version of the Truth Ratio, which we would still report.
Proposed Experiments
The main testbed is TOFU Forget10 on Llama2-7b, chosen because it has a retain-only reference model and a published baseline for WAGLE. We run all the unlearning methods above, each at five strength levels, with 200 forget and 200 retain questions per checkpoint. The target and ordinary peers use the unlearned model. The stubborn peer is a separate model under Belief Engine control.
A second testbed is WMDP-Bio with RMU on Zephyr-7B. There the peer argues for a wrong option or the correct option in multiple-choice form. With no retrained reference, the null is set by the base model's behaviour on questions it answers at random-chance level.
We report four quantities: $\Delta$ with bootstrap confidence intervals, re-elicitation rate beyond the retain-only reference, budget-matched leak@k, and retain utility.
The first decisive check is whether the retain-only model's $\Delta$ is near zero while NPO's and WAGLE's are clearly positive. If the reference itself shows asymmetry, the perturbed answers are not truly matched and the design must be repaired before anything else.
The second is whether $\Delta$ rises monotonically as unlearning is weakened to preserve utility, and correlates with the linear-probe accuracy.
The third is whether RULE-GradDiff, near zero on leak@k, still shows a positive $\Delta$.
The fourth is whether the no-stubbornness and single-turn ablations shrink $\Delta$. If they do, the stubborn-peer mechanism, not just in-context evidence, is doing the work.
A negative result on the third check combined with a positive result on the first two would still be informative. It would mean sampling-hardened unlearning also resists social pressure, and $\Delta$ would become a black-box certificate of forgetting rather than an attack.
References
- Multi-Agent Systems are Mixtures of Experts: Who Becomes an Influencer?
- Biases in Opinion Dynamics in Multi-Agent Systems of Large Language Models: A Case Study on Funding Allocation
- Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding
- An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations
- The social system identification problem
- Active Sensing of Social Networks
- Agents Are All You Need for LLM Unlearning
- Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation
- The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning
- Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
- WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models
- TOFU: A Task of Fictitious Unlearning for LLMs