Teaching Domain-Adapted Chat Models to Accept Retrieved Updates

Specialist chat assistants in medicine or law are usually built in stages. A general model is first trained further on raw domain text, then taught on domain conversations, and then connected to a search system. That search system, called retrieval-augmented generation, looks up documents and pastes them into the prompt so the model can use current guidelines. The stages pull against each other. Each round of domain training writes more facts into the model's weights, which makes the model more likely to ignore a retrieved document that says those facts have changed. Methods that force the model to follow the document are tested only on short-answer questions. In a conversation, they may make the model paste chunks of text or change topic abruptly.

This note proposes to measure how resistance to retrieved updates builds up across the stages of domain adaptation. It then tests a specific claim about when to fix it. The claim is that a small amount of practice at following updates, written as natural dialogue and mixed into the adaptation stages while the model's internal fact-storing pathways are still forming, gives more faithful and more fluent updates than the same practice applied after adaptation is finished. The same practice is meant to avoid making the model believe whatever it retrieves.

Core Claim

Two findings in the literature point to a mechanism that nobody has connected. The knowledge-circuits study of continued pretraining (How Do LLMs Acquire New Knowledge?) traced the small sets of internal connections that store individual facts. It found that these pathways first rearrange quickly (a formation phase) and then settle into a fixed layout whose weights keep strengthening (an optimization phase). It also found that new facts tied to entities the model already knows are absorbed faster. Separately, conflict benchmarks show that strong stored beliefs cause the model to reject contradicting context. FaithEval reports GPT-4o falling to 47.5% accuracy on counterfactual science contexts, and Wang et al. tie this entity bias to how often an entity appeared in pretraining.

We hypothesize three things. First, resistance to retrieved updates rises sharply at the switch from formation to optimization for the domain facts. Second, it is worst for facts anchored to familiar entities. Third, update-following behavior is cheaper to install before that switch than after it. The conversational part of the claim is that the practice data must itself be dialogue. We expect question-answer style conflict training, as in CLEAR, to restore faithfulness but to increase verbatim copying in multi-turn chat. DeepCopy documented this copying failure as a harm to perceived coherence.

Research Program

The first component is a controlled testbed, because real domain corpora cannot tell us which facts a model holds or how strongly. We add a set of synthetic but plausible domain facts to a domain continued-pretraining corpus. Examples are dosing thresholds, contraindications, and procedural rules. Half of the facts attach to well-known entities and half to invented ones, matching the relevant-versus-isolated split in the circuits study. We then write retrieval documents that revise a subset of these facts, and we build multi-turn patient-style conversations on the style of HealthCareMagic-100k, as used by ChatDoctor. Each conversation raises a revised fact in the middle and returns to it several turns later.

Following the defensive view in SeCon-RAG, where documents that clash with the model's beliefs are filtered out, we also include poisoned documents. These state false revisions and lack the dated, authoritative provenance that genuine updates carry. A model that simply believes all context therefore scores poorly. Provenance is the only signal that separates the two cases. We state this assumption openly rather than claim to resolve SeCon-RAG's opposing definition of trust in general.

The second component is a stage-by-stage diagnosis. We follow the ChatDoctor and MoZi-style curriculum (MoZIP): base model, then general instruction tuning, then domain continued pretraining, then domain dialogue tuning. We save checkpoints densely during continued pretraining. At each checkpoint we measure three things: how often the model adopts retrieved updates, how circuit layouts for the synthetic facts shift (using the knowledge-circuits study's attribution method, with overlap between successive layouts as the stability signal), and how fluent the model is.

The third component is the intervention, which we call interleaved update practice. A few thousand dialogue examples are mixed into the continued-pretraining and dialogue-tuning stages. In each example, a retrieved dated update contradicts a trained fact and the reference reply weaves the new fact in naturally. Some examples instead carry an unsourced contradiction, and the reference reply keeps the prior belief and flags the discrepancy. The key comparison places the same examples, with the same count and compute, at different points in the curriculum.

Baselines

The naive reference point is the fully adapted model with retrieval on every turn and a system prompt telling it to rely strictly on the retrieved documents. This is the setup most deployed domain assistants use. FaithEval shows prompting of this kind is weak, so it sets the floor.

A second baseline adds RAGate, a learned gate that decides per turn whether to retrieve. On KETOD it retrieved on about 16% of turns while keeping BLEU/ROUGE comparable and raising confidence from 8.36 to 9.30. It is the strongest fluency-preserving competitor, but it has no conflict handling. We retrain its gate on our dialogues.

The strongest faithfulness competitor is CLEAR, which applied after the full curriculum. CLEAR uses a hidden-state probe to detect conflict and then conflict-aware fine-tuning with a loss that guides attention toward the evidence. Its reported result is 74.4% F1 on FaithEval with LLaMA-3.1-8B, against 71.6% for Context-DPO. We run it on the same base model family, with a tuning budget in examples and GPU-hours matched to our intervention. We also report FaithEval, so the comparison against its published number is direct.

We also include ContextFocus, which adds a steering vector to intermediate-layer activations at inference time and needs no fine-tuning, extracted from 1,500 examples. We found no reported numerical result on Lacuna for ContextFocus on ConFiQA or NQ-SWAP, so we reproduce it and tune its multiplier on held-out dialogues.

The ablation that isolates our key ingredient is post-hoc update practice. It uses exactly our dialogue examples, but as a final fine-tuning stage after adaptation. A second ablation keeps our placement but converts the examples to question-answer form.

Beating the field means a specific pattern on held-out conversations. Our method must achieve higher update adoption and lower reversion to the stale fact in later turns than CLEAR and post-hoc practice. It must reach poison adoption no worse than the naive prompt, and coherence ratings plus copy rates no worse than RAGate. It must also stay within one point of the unmodified adapted model on domain task quality.

Why not just…?

Why not just tell the model to trust the documents? FaithEval reports that instruction-following models still collapse on counterfactual contexts. It also found that prompts urging vigilance produce sycophantic false "unanswerable" claims; Phi-3-medium fell to 7.4% on that task. We keep this as the naive baseline and expect it to fail most on facts tied to familiar entities.

Why not just apply CLEAR or steering once adaptation is done? This is the main threat, and the post-hoc ablation exists to settle it. If post-hoc practice with our dialogue data matches interleaved practice on adoption, fluency, and poison resistance at equal budget, then the timing claim is false. The contribution would then shrink to the diagnosis and the conversational benchmark, and we would report it that way. Our prior against post-hoc repair comes from the circuits finding that later training only refines a fixed layout, which suggests late corrections sit on top of stable pathways rather than rewiring them. That is an inference to be tested, not an established result.

Why not skip weight-based domain training and rely on retrieval alone, as RADA does for classification? RADA replaces domain pretraining with a plug-in retrieval interface and gains +1.59 points over fine-tuning. However, it was tested only on BERT-scale discriminative tasks. Don't Stop Pretraining and the staged medical models suggest weight-based adaptation carries vocabulary and reasoning that retrieval alone does not. We include a retrieval-only arm, the instruction-tuned model with no domain stages, to measure what it costs in domain dialogue quality.

Why not reason through the documents with explicit notes? Chain-of-Note improves robustness to noisy retrieval, but explicit notes took 12.0 s per query against 0.6 s, which is too slow for live dialogue. It was also evaluated only on question answering. We report it only as a latency-unconstrained reference on a subset.

Proposed Experiments

The testbed uses a 7–8B open model through the four-stage curriculum. Evaluation covers three settings. The first is our synthetic-update dialogues, with a held-out split of revised facts never seen in any practice data. The second is FaithEval and ConFiQA, to check that gains transfer to standard conflict benchmarks. The third is KETOD, where joint goal accuracy (SimpleToDPlus reported 28.6%) checks that task tracking survives, and the Document Grounded Conversations set, where ITE+DD reported coherence 0.90 against 0.56 under the same human protocol we reuse.

Faithfulness metrics are update adoption, later-turn reversion to the stale fact, and poison adoption. Fluency is measured by blinded human coherence and naturalness ratings, plus a copying score: the longest overlap between the reply and the retrieved passage.

Three results would be decisive. The first is whether update adoption falls most steeply at the checkpoint where circuit-layout overlap for the synthetic facts stops changing, and whether the drop is larger for facts tied to familiar entities. If adoption instead declines smoothly with training steps, the mechanistic story fails. The second is whether interleaved placement beats post-hoc placement at matched budget on held-out facts. The third is whether dialogue-form practice lowers copying relative to question-answer-form practice at equal adoption. A null result on the second check reduces the proposal to a well-controlled benchmark. Null results on both the second and third checks mean the simpler recipe of running CLEAR after adaptation is the right answer, and we would say so.