A retrieval-augmented language model answers questions using documents pulled into its prompt. When a document contradicts what the model learned in training, the model must choose: trust the document, trust its memory, or decline to answer. Current methods make this choice one prompt at a time, as if each document came from nowhere. In practice, documents come from sources such as a website, a tool, a user, or another agent, and a model often hears from the same source many times. Some sources are honest but stale. Some are sloppy and wrong at random. Some are strategic: they tell the truth on facts the model can check, which builds trust, and lie on facts it cannot check. A model that judges a source by how often it agreed on verifiable facts will be fooled by the third kind exactly where it is most exposed.
We propose to give the model a small, explicit belief about what kind of source it is talking to, updated turn by turn, and to let that belief drive a calibrated abstention rule. The key ingredient comes from interactive partially observable Markov decision processes (I-POMDPs), which are planning models in which an agent's hidden state includes other agents' goals and beliefs. Here that means one hypothesis about the source is "this source knows what I don't know, and lies there." In two sentences: a strategic source's errors are concentrated on the model's own blind spots, so the model's estimate of a source's reliability should depend on how confident the model itself is on each question. Combining that estimate with a risk-controlled abstention threshold should keep error rates bounded against trust-building adversaries, without the over-suspicion that sophisticated reasoners fall into against harmless noise.
Core Claim
Existing conflict-resolution methods treat context as a passive, per-prompt input. Chain-of-Note writes reading notes and abstains when documents are useless. CLEAR probes hidden states for latent conflict and then trains the model to follow context. ContextFocus and PH3 steer activations or prune attention heads toward context or memory, but a human must say which one to favour. Critic-Reasoner masks conflicting spans. None of these keeps memory of a source across turns. KnowGuard is multi-round and abstains to ask follow-up questions, but it adjudicates against a fixed, authoritative knowledge graph and does not model the dialogue partner's intent.
The obvious fix for missing memory is a running reliability score per source, for example a count of past disagreements. This is the main competitor, and the proposal is built around beating it on a specific, falsifiable point. A running score assumes a source's error rate is the same on questions the model can verify and questions it cannot. That assumption holds for honest and randomly noisy sources. It fails for strategic sources. Our claim is that a source model in which the error probability depends on the model's own confidence matches the running score on non-strategic sources and keeps selective risk at its target on strategic ones, where the running score's risk rises sharply on low-confidence questions. The two parts of that claim matter equally. The second shows the intentional model is worth having. The first guards against the "Machiavellian trap" documented in (Mal)adaptive Mentalizing, where an over-sophisticated agent misreads simple behaviour as manipulation.
Research Program
The work has three pieces: a benchmark, a belief filter, and a selection rule.
The benchmark. No existing benchmark has repeated interaction with a single source of controlled type, so we build episodes from established datasets. Each episode pairs one source with a sequence of 20–40 questions. An honest source supplies gold passages, including post-cutoff passages from RealTimeQA that legitimately conflict with stale parametric memory; this tests whether the method still accepts genuine updates and avoids the stubbornness reported by Wang et al.. A noisy source mixes in irrelevant retrieved documents and ConFiQA counterfactuals at random positions. A strategic source is white-box: it estimates the target model's confidence on each question by sampling-consistency on the target itself, answers truthfully on high-confidence questions, and injects ConFiQA counterfactuals or SciAux-style misleading snippets on low-confidence ones. An adaptive variant lowers its lie rate when the target starts abstaining. Question pools come from NQ, ConFiQA (single-hop QA, multi-hop with one counterfactual, and multi-hop with several), RealTimeQA, and SciAux from Thinking in a Crowd. Mixed episodes, in which the source type switches partway through, test recovery speed.
The belief filter. At each turn $t$ the model computes three signals. The first is its parametric answer and confidence $c_t$, estimated by agreement across sampled answers without context. The second is whether the context's answer disagrees with the parametric one. The third is a latent-conflict probe score trained as in CLEAR. A particle filter in the style of the interactive particle filter holds hypotheses over the source type $\tau$ (honest, noisy, or strategic) and that type's parameters. For the strategic type, the probability that the passage is false on turn $t$ is
$$P(\text{false}_t \mid \tau=\text{strategic}, c_t) = \sigma(\alpha - \beta\, c_t),$$
where $\sigma$ is the logistic function and $\alpha, \beta$ are per-particle parameters with $\beta > 0$. In words, the source lies more where the model knows less. This term is the depth-one mentalizing step: the agent models the source's model of the agent. Honest and noisy types use error rates that do not depend on $c_t$. Honest sources err rarely, and when they disagree with memory it is mostly because memory is wrong, which lets RealTimeQA updates through. Only the honest type uses the fact that the model's own wrong answers explain disagreement. This mirrors the finding in (Mal)adaptive Mentalizing that explicit noise hypotheses "explain away" deviations and reduce paranoia. Ground truth is never fed to the filter. Updates use only the signals above, plus occasional verifiable turns on which parametric confidence is high.
The selection rule. The model has three actions: answer from context, answer from memory, or abstain. Let $b_t$ be the current belief over source types, so that $\hat r^{\text{ctx}}_t = \mathbb{E}_{b_t}[P(\text{false}_t \mid \tau, c_t)]$ is the estimated risk of following context. The estimated risk of following memory is $\hat r^{\text{mem}}_t = 1-\tilde c_t$, with $\tilde c_t$ the confidence after isotonic calibration. The model takes the lower-risk answer and abstains when
$$\min(\hat r^{\text{ctx}}_t, \hat r^{\text{mem}}_t) > \lambda,$$
with $\lambda$ chosen on held-out episodes so that risk on answered questions stays below a target $\epsilon^*$. This is the risk-coverage formulation of Selective Classification for Deep Neural Networks, applied at the episode level. Because the turns within an episode are not exchangeable, the calibration resamples whole episodes. We report the resulting guarantee as an empirical check, not a theorem. When the model answers from context, the answer can optionally be produced with ContextFocus steering. When it answers from memory, the context is dropped, which avoids the rationalization failure reported for reasoning models.
Baselines
The naive reference point is the base model (Llama-3.1-8B, the model CLEAR uses, and Qwen3-8B with thinking on and off, as in SciAux) reading every passage with no abstention. A second naive point is "always trust memory when it conflicts with context," which bounds how much is lost to stubbornness on honest-update episodes.
The strongest per-turn method is CLEAR's conflict probe combined with a Chain-of-Note abstention prompt, swept over a confidence threshold to trace a risk-coverage curve. On published benchmarks, CLEAR reports 74.4% F1 on FaithEval with Llama-3.1-8B against a strongest prior baseline of 71.6% (CLEAR), and Chain-of-Note reports +10.5 correct rejection on RealTimeQA (Chain-of-Note). We will reproduce both numbers before using these methods on our episodes. We also run Distraction-aware Answer Selection, which scores documents by query likelihood, and Self-Knowledge Guided Retrieval, which gates retrieval by nearest-neighbour self-knowledge, as per-turn selectors. The strongest conflict-specific system is the Critic-Reasoner, with its reported 25.99% accuracy gain in adversarial conflict settings. It has no abstention, so we add the same calibrated threshold on its output confidence.
The key ablation is the running reliability score: a Beta-Bernoulli posterior on each source's disagreement rate, fed the same signals and calibrated with the same selection rule. It is our filter with $\beta$ fixed at zero, so any gap between the two isolates the intentional term. Further ablations remove memory entirely (the filter is reset every turn), use only a strategic hypothesis with no noise type (to provoke paranoia), and swap the probe for sampling-consistency alone.
For fairness, every method sees the same retrieved passages and the same base model. Every method gets the same calibration episodes and a search over 20 threshold values. Learned probes use the same 1.5k training examples that ContextFocus reports needing. The benchmark is new, so there is no reported result found on Lacuna for any method on these multi-turn source-typed episodes, and all comparisons there are our own runs. Beating the running score means two things. On strategic episodes, our coverage at $\epsilon^* = 5\%$ must be significantly higher, or our realized risk must stay within target where the running score's does not. On honest and noisy episodes, we must be no worse, which we define as a coverage gap inside the seed-to-seed confidence interval.
Why not just…?
Why not just keep a running reliability score per source? This is the real threat, and the proposal treats it as the main ablation, not a straw man. On honest and noisy sources we expect the two to tie. If they also tie on strategic sources, then knowledge-conditioned lying is either not learnable from these signals or not damaging in practice, and the intentional component should be dropped. The trust-building adversary is designed so that the two predictions diverge. On that adversary the running score is pushed toward "reliable" by verifiable turns. We will report risk separately by bins of the model's own confidence to show where the divergence occurs.
Why not just ask the model, or a reasoning model with the full history in context, whether to trust the source? Direct self-assessment was the least effective gating strategy in Self-Knowledge Guided Retrieval. On SciAux, two misleading snippets drove Qwen3-8B with thinking on to 77.03%, against 96.22% with no context and 74.87% with thinking off (Thinking in a Crowd). Long-context ingestion also raised overconfidence in KnowGuard. Still, a reasoning model given the whole episode transcript might infer the adversary's pattern. We therefore include it as a baseline with an explicit prompt describing the three source types. This is a fair test of whether the explicit filter adds anything beyond describing the hypotheses in words.
Why not just distrust any context that conflicts with memory? It blocks legitimate updates, and models already show both stubbornness on frequent entities (Song et al.) and gullibility (Hong et al.). The RealTimeQA honest-update episodes measure this directly. "Always trust memory" should fail there, and our method must accept those updates at a rate close to a per-turn context-following method.
Proposed Experiments and Decisive Checks
The main result is a set of risk-coverage curves and coverage at fixed risk ($\epsilon^* \in \{2\%, 5\%, 10\%\}$), broken down by source type and dataset, averaged over five seeds and two base models.
Three secondary measures follow the I-POMDP literature. The first is the false-suspicion rate, defined as abstention against honest and noisy sources beyond what the running score produces. The second is the number of turns needed to identify a source type, which parallels the five-versus-two trials contrast reported for over- and well-calibrated mentalizers ((Mal)adaptive Mentalizing). The third is recovery time after a mid-episode type switch.
Three checks decide the outcome. First, on strategic episodes, does the full filter hold realized risk at target where the running score exceeds it, specifically in the lowest-confidence bin? Second, on noisy episodes, does the strategic-only ablation become paranoid while the full mixture does not? This would reproduce, in a language-model setting, the noise-mitigates-paranoia effect. Third, against the adaptive adversary, which learns to lie less once it sees abstentions, does the risk bound still hold? A failure there would show that the depth-one model needs a deeper recursive level, which would be the natural follow-up.
A negative result is also informative. If parametric confidence $c_t$ is too poorly estimated for the knowledge-conditioned likelihood to separate source types, then the check exposes a limit of internal self-knowledge signals, which is the central tension between probing methods and external-knowledge methods such as KnowGuard.
References
- Chain-of-Note: Enhancing Robustness in Retrieval-Augmented Language Models
- Probing Latent Knowledge Conflict for Faithful Retrieval-Augmented Generation
- ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models
- Cutting Off the Head Ends the Conflict: A Mechanism for Interpreting and Mitigating Knowledge Conflicts in Language Models
- From Extraction to Deduction: Resolving Functional Misalignment in RAG via a Collaborative Critic-Reasoner Framework
- KnowGuard: Knowledge-Driven Abstention for Multi-Round Clinical Reasoning
- (Mal)adaptive Mentalizing in the Cognitive Hierarchy, and Its Link to Paranoia
- A Causal View of Entity Bias in (Large) Language Models
- Thinking in a Crowd: How Auxiliary Information Shapes LLM Reasoning
- Monte Carlo Sampling Methods for Approximating Interactive POMDPs
- Selective Classification for Deep Neural Networks
- Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering
- Self-Knowledge Guided Retrieval Augmentation for Large Language Models
- Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse
- Why So Gullible? Enhancing the Robustness of Retrieval-Augmented Models against Counterfactual Noise