Informative Abstention from Knowledge-Ablated Synthetic Dialogues

When a chat assistant is asked about something its sources do not cover, it usually does one of two things. It invents a plausible answer, or it says "I don't know" and stops. The first is a hallucination. The second looks honest, but people rate it as unhelpful. Qian et al.'s BUSY pipeline found exactly this: its database-grounded shopping dialogues refused to invent missing product specifications, and human raters judged them less informative than ungrounded dialogues that made the specifications up. This note argues that the "honesty penalty" is mostly a property of the training data, not a law of the tradeoff. The training dialogues model a missing fact as a bare refusal, when they could model it as a refusal plus the closest things the assistant does know, plus a useful next step.

The proposal is to generate synthetic multi-turn dialogues in which we deliberately delete specific facts from the assistant's source material while a simulated user, who can see those facts, asks about them, presses for them, or asserts wrong values for them. Because we removed the facts ourselves, we know exactly which turns sit on the edge of the assistant's knowledge. We can then label, without a judge model, which candidate replies fabricate, which refuse emptily, and which refuse informatively. Training on these labeled turns, conditioned on the full conversation history, should give an assistant that rarely fabricates on deleted facts across ten turns, holds that line under user pressure, and still receives informativeness ratings close to an ungrounded model.

Core Claim

The claim has two parts. The first is about labels. Counterfactual deletion of knowledge turns the hard question "is this a truthfulness-helpfulness conflict?" into something known by construction. CAMP showed that the two goals conflict in only about 11% of preference pairs, and that treating those pairs differently helps. However, CAMP finds conflicts with GPT-4 scoring on single-turn UltraFeedback data. In our generator every question about a deleted fact is a conflict instance, every question about a retained fact is a synergy instance, and the gold value of each deleted fact is available for checking.

The second part is about targets. The preferred reply at a conflict turn is neither silence nor a guess. It declines the deleted fact explicitly, supplies the most relevant retained facts, and offers a clarifying question or an alternative. Informativeness is scored so that padding cannot inflate it. We predict that this data shifts the hallucination–informativeness frontier outward relative to bare-refusal training of the same size. It should not simply move a model to a different point on the same curve.

We restrict the claim to faithfulness hallucination, meaning claims that go beyond the provided context, in the sense defined by He et al.. We do not attempt to adjudicate the model's parametric knowledge. This restriction is what makes the labels exact. We still record whether a fabricated value happens to match the deleted gold value, which separates lucky recall from grounded answering.

Research Program

The generator starts from two kinds of grounding: Wikipedia passages behind FaithDial topics, and structured product records in the style of ShopDial. For each dialogue we sample a deletion set of facts, typically 20–40% of the attributes or sentences. The assistant agent sees only the remaining knowledge.

The user simulator is a separate agent that sees the full knowledge and a goal. The two-agent design follows the finding in the multi-agent medical and self-chat work (for example Bao et al.) that a single prompt writing both sides produces an overly cooperative user. We want real friction. The simulator's goals mix four kinds of request: ordinary questions about retained facts; questions about deleted facts; follow-up pressure after an abstention ("surely you can estimate it"); and false-premise assertions that state a wrong value for a deleted fact. That last type extends Pressure-Tune, which only tested pressure against facts the model actually knows, into the case where the model lacks the fact. Sessions are written as whole conversations rather than stitched turns, following Bae et al. and TurnWise, because turn-by-turn construction loses cross-turn dependencies.

At each assistant turn we sample several candidate replies from differently prompted assistants. We label them by construction. Any reply that asserts a value for a deleted fact counts as a fabrication. We check this with NLI (natural-language inference, a model that judges whether one text supports another) against the visible context and against the deleted fact's slot. A reply that declines and adds nothing counts as a bare refusal. A reply that declines, adds retained facts relevant to the question, and proposes a next step counts as informative abstention.

To score informativeness we count supported claims only after CORE's filtering, which removes redundant and bleached sub-claims. This stops a reply from buying informativeness with generic filler. CORE showed that raw FActScore rated uninformative Mistral-7B text at 83.0% but only 36.2% once filtered.

A fraction of histories are "poisoned": we inject an earlier assistant turn that fabricated a deleted fact. In those histories the preferred reply retracts the fabrication. This targets the compounding problem, in which an early unsupported claim hardens into accepted background for later turns.

Training is turn-level preference optimization conditioned on the full history. The preference order is informative abstention above bare refusal above fabrication at conflict turns. At synergy turns, a correct, detailed answer is preferred over an unwarranted refusal, which penalizes over-refusal of the kind GRAIT documents. We use CAMP's adaptive margin on the conflict turns. Here the margin is switched by the construction label, and no GPT-4 conflict score is involved.

Baselines

The naive reference point is the base instruction-tuned model, Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct (the models CAMP used), placed as the assistant in the same held-out knowledge-ablated dialogues with only a standard grounding instruction. Every method uses the same held-out dialogues, deletion sets, and user-simulator scripts, so turn-by-turn comparisons are paired.

The strongest prompting competitor is the same base model with a carefully engineered system prompt. The prompt explicitly asks it to decline facts missing from the context, give related known facts, and ask clarifying questions. We will tune the prompt on a development split with the same number of iterations we spend tuning our data mixture. Beating it means lower fabrication on deleted facts under pressure and on poisoned histories, with human informativeness no worse.

The strongest data competitor is BUSY-style training (Qian et al.): the same generator and the same number of dialogues, with bare "I don't know" as the target at deletion turns. This is also the ablation that isolates our key ingredient, the informative-abstention target. For reference, BUSY reports LLM-judged truthfulness of 4.70 versus 4.55 for PLACES on ShopDial. Its informativeness deficit is described qualitatively, and no reported numeric result was found on Lacuna for BUSY's human informativeness gap. Beating it means matching its fabrication rate while closing most of the informativeness gap to the ungrounded top-down generator, which we also run as the high-informativeness reference.

The strongest alignment competitor is CAMP trained as published on single-turn UltraFeedback, with the same number of optimizer steps. CAMP reports a HaluEval hallucination rate of 3.20% against 4.35% for standard DPO. We will reproduce this number as a sanity check before comparing on multi-turn metrics, where no CAMP result exists.

The strongest inference-time competitor is certainty-based response ranking (CRR) from Wan et al., which samples N replies and keeps the most certain. It reports 98.6% faithfulness on FaithDial with OpenLlama-3B, nucleus sampling, and P-CRR. We add a stronger variant that reranks with our own labeler (VISTA-style support minus fabrication, plus CORE-filtered informativeness). Both run at N=8 on the base model, and we report their extra compute.

Further ablations isolate the remaining design choices. One removes pressure turns. One removes poisoned histories. One flattens the same preference pairs into single-turn examples with no history. One replaces construction labels with GPT-4 judge labels. The last isolates the value of knowing the conflict set exactly.

Why not just…?

Why not just prompt the model to give related facts when it doesn't know? This is the most serious threat. A capable model may produce informative abstentions on a quiet first turn from instructions alone. The evidence that prompting breaks down comes from harder conditions. In Pressure-Tune, Llama-3.1-8B resisted sycophantic pushback only 6.40% of the time before training. GRAIT reports that prompting-based refusal causes over-refusal on known questions. Neither result covers our exact setting, so our experiments settle it. We compare the tuned prompt and our model separately on first-ask turns, pressure turns, false-premise turns, and poisoned histories, all at turns 1 through 10. If the prompt matches us on every slice, the data contribution is not needed, and we will say so.

Why not just apply CAMP or DPO to single-turn data? CAMP's conflict label ignores conversation history. In dialogue, whether a reply fabricates can depend on what earlier turns established, and on whether an earlier turn already contained a fabrication. The flattened single-turn ablation tests this directly. The same pairs without history should do worse on poisoned-history retraction and on late-turn fabrication, because TurnWise found that multi-turn quality decays unless training is multi-turn.

Why not just sample many replies and rerank at inference time? Certainty ranking chooses what the model is most sure of, and a bland refusal is often the most probable sequence. Wan et al. note this ranking has no informativeness term. Our informativeness-aware reranker removes that objection. The remaining question is whether reranking can recover behavior the base model rarely produces, such as retracting its own earlier claim. We will report whether the N=8 reranker matches the trained model, and at what multiple of the trained model's inference cost.

Proposed Experiments

The primary evaluation uses held-out knowledge-ablated dialogues from both domains. We plot fabrication rate on deleted facts against CORE-filtered supported claims per turn, broken out by turn index and by request type. The decisive result is a frontier shift: at equal or lower fabrication than the bare-refusal baseline, more informative supported claims, and blind human informativeness ratings statistically indistinguishable from the ungrounded top-down generator. We also report over-refusal, meaning the rate of declining facts that are present in context, which must stay near the base model's rate.

Transfer tests use existing benchmarks. On the FaithDial test set we measure faithfulness against the 98.6% CRR figure and the 77.4% human-rated faithfulness of FaithDial's T5 model. On the VISTA 140-dialogue benchmark we use VISTA's claim categories, which separate abstention from lacking evidence and contradiction. VISTA identified abstentions correctly in 90.6% of CoCoNot cases, so we trust its abstention label. On CoCoNot we measure appropriate abstention. On TurnWiseEval we check that general multi-turn ability does not regress. Parametric-leakage analysis compares fabricated values against the deleted gold values.

The finding that would most change our view is if the informative-abstention model fabricates more on pressure turns than the bare-refusal model. That would mean giving adjacent facts invites the user to push for more, and that the honest path needs a different target.