💬 NLP

Clarify, Abstain or Answer? Strategising in Conversation with Belief-Augmented Generation

This paper proposes Belief-Augmented Generation (BAG), a framework that leverages an LLM's own belief state derived from multiple sampled responses to autonomously decide whether to answer, clarify, or abstain in ambiguous conversations, thereby improving accuracy and strategy faithfulness compared to standard prompting.

Original authors: Joris Baan, Wilker Aziz, Barbara Plank, Raquel Fernández

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Joris Baan, Wilker Aziz, Barbara Plank, Raquel Fernández

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Core Problem: The "Overconfident" AI

Imagine you ask a large language model (LLM) a tricky question, like: "Who sang the female part on 'Gimme Shelter'?"

If you ask a standard AI, it will usually just pick one answer and say it with total confidence, even if it's guessing. It might say, "It was Merry Clayton," without realizing that if you meant the live tour version, the answer is actually "Lisa Fischer." The AI doesn't know it's unsure; it just acts like a know-it-all.

The researchers found that these models rarely say, "I'm not sure which one you mean," or "I don't know the answer." They just guess.

The Solution: The "Belief-Augmented Generation" (BAG)

The authors propose a new method called Belief-Augmented Generation (BAG). Think of this as giving the AI a "thinking room" before it speaks.

Instead of generating just one answer, the AI is asked to generate 10 different possible answers to the same question first. It then looks at this list of 10 answers (which the paper calls its "Belief State") and asks itself: "What do these 10 answers tell me about how sure I am?"

Based on what it sees in that list, the AI chooses one of three strategies:

Answer Directly: If all 10 answers are basically the same (e.g., all 10 say "Merry Clayton"), the AI knows it's confident. It gives a direct answer.
- Analogy: It's like a jury where 10 jurors all agree on the verdict. The foreman just announces the result.
Clarify: If the 10 answers are split into two distinct groups (e.g., 5 say "Merry Clayton" and 5 say "Lisa Fischer"), the AI realizes the question is ambiguous. Instead of guessing, it asks the user for clarification.
- Analogy: It's like a waiter who hears a table order "The special" but sees two different specials on the menu. Instead of guessing, the waiter asks, "Did you mean the steak or the fish?"
Abstain (Say "I Don't Know"): If the 10 answers are all over the place and contradict each other wildly (e.g., one says "Merry Clayton," another says "The Beatles," another says "A robot"), the AI realizes it doesn't know the facts. It politely declines to answer.
- Analogy: It's like a detective who finds 10 different suspects with no evidence linking any of them to the crime. The detective admits, "I can't solve this yet," rather than arresting a random person.

How They Tested It

The researchers tested this on a dataset of "Ambiguous Questions" (questions that could have multiple meanings). They compared:

Standard AI: Just answers directly.
Prompt-Only AI: Told to "be careful" but doesn't see its own list of guesses.
BAG AI: Sees its own list of 10 guesses and reasons through them.

The Results

Better Accuracy: The BAG method improved the accuracy of the answers across six different AI models.
Smarter Choices: The BAG AI was much better at knowing when to ask a question and when to admit it didn't know. The standard AI almost never asked for help or admitted ignorance.
Faithfulness: The BAG AI's choices matched its internal "confidence" (the diversity of its 10 guesses) much better than the other methods. If it was confused, it acted confused; if it was sure, it acted sure.

The Catch

While the system works well, it is still hard for the AI to perfectly tell the difference between:

Real Ambiguity: "I don't know if you mean the movie or the book." (Needs a Clarification).
Hallucination: "I'm making up facts because I don't know." (Needs an Abstain).

Sometimes the AI thinks a question is ambiguous when it's actually just making things up, or vice versa. But overall, giving the AI a chance to "look at its own work" before speaking makes it a much more honest and helpful conversational partner.

Summary

The paper introduces a way to make AI models less overconfident. By forcing them to generate multiple answers first and then analyze the differences, the models learn to Clarify when they are confused, Abstain when they are clueless, and Answer only when they are sure. This makes them better at handling tricky, real-world questions.

Technical Summary: Clarify, Abstain or Answer? Strategising in Conversation with Belief-Augmented Generation

Problem Statement

Large Language Models (LLMs) define a probabilistic distribution over text, which serves as a representation of uncertainty or "belief." While existing research leverages this distribution for narrow tasks like decoding (e.g., Minimum Bayes Risk) or selective prediction (e.g., detecting hallucinations via semantic entropy), these approaches typically require manual interventions, external decoders, or fine-tuning. Crucially, standard LLMs lack a mechanism to "introspect" or directly exploit their own probabilistic uncertainty during generation. Empirical evidence suggests that without explicit guidance, LLMs rarely clarify ambiguous inputs or abstain from answering when uncertain; instead, they default to direct answers, often ignoring input ambiguity or internal factual inconsistencies. This leads to hallucinations or incorrect answers in multi-turn ambiguous Question Answering (QA) settings.

Methodology: Belief-Augmented Generation (BAG)

The authors propose Belief-Augmented Generation (BAG), a training-free framework that grounds LLMs in their own belief state to enable richer conversational strategies: answering directly, asking for clarification, or abstaining.

Core Mechanism

BAG operates as a two-stage process:

Belief State Construction: Given an input $x$ , the model generates $K$ samples ( $Y^{(1)}, \dots, Y^{(K)}$ ) via Monte-Carlo sampling. Unlike methods that reduce these samples to a single metric (e.g., entropy or majority vote), BAG retains the full set of $K$ textual responses as the belief state $B_\theta(x)$ . This state captures not only the diversity of plausible answers but also linguistic cues such as hedging, refusals, or contextual variations.
Strategy Reasoning: The model is prompted with the original user question and the $K$ $K$ candidate answers (the belief state). It is instructed to analyze this state and select one of three strategies:
- Direct Answer: If the belief state shows high consistency (low uncertainty) regarding the answer.
- Clarification: If the belief state contains semantically distinct but plausible answers corresponding to different interpretations of the question (data uncertainty/ambiguity).
- Abstain: If the belief state contains contradictory facts indicating a lack of knowledge (epistemic uncertainty/hallucination) where no useful clarification exists.

Prompt Variants

The authors explore three prompt variants (BAG1, BAG2, BAG3) with increasing complexity in instructions:

BAG1: Basic consistency checking.
BAG2: Incorporates uncertainty-specific language and distinguishes between ambiguity (clarify) and hallucination (abstain).
BAG3: Requires the model to cluster generations by meaning, interpret the clusters, and then decide on a strategy.
BAG+: An extension where, if the model chooses to clarify, it simulates a user response (using an auxiliary LLM) and re-evaluates the belief state to decide on a final answer or further abstention.

Key Contributions

Direct Access to Belief States: The paper introduces a method to feed the raw textual representation of a model's uncertainty directly into the prompt, allowing the model to reason over its own distribution without external tools or training.
Strategy Selection: Unlike prior work focused on selecting a single generation or hedging a claim, BAG enables the model to dynamically choose between answering, clarifying, or abstaining based on the nature of its uncertainty.
Training-Free Applicability: The approach requires no fine-tuning, external data, or access to model weights, making it applicable to both open and closed models.
Empirical Analysis of Routing: The study provides a detailed decomposition of how models route questions to different strategies and evaluates the faithfulness of these decisions to the underlying belief state entropy.

Results

Experiments were conducted on the AmbigQA dataset (1,832 ambiguous questions) across six models (OLMo2/3, Qwen3, Gemini-2.5-Flash).

Performance Gains: BAG significantly improves QA accuracy compared to standard direct generation and prompt-only baselines (Strategy-Augmented Generation, SAG). In many cases, BAG+ outperforms even the "disambiguation oracle" (which provides the correct intent), demonstrating that the model can successfully identify when it lacks factual knowledge and abstain rather than guessing.
Faithfulness: BAG's strategic decisions are more faithful to the model's internal belief state entropy than SAG. As belief state entropy increases, BAG models increasingly choose to clarify or abstain, whereas SAG and direct generation rarely do so.
Strategy Decomposition: The accuracy gains are driven by different mechanisms depending on the model. For instance, for Qwen models, the improvement is primarily driven by the abstain strategy (avoiding hallucinations), while for Gemini, the contributions of clarification and direct answering are more balanced.
Challenges in Disentanglement: While BAG improves performance, distinguishing between interpretation variation (requiring clarification) and hallucination variation (requiring abstention) remains challenging. Models sometimes generate "perceived ambiguity" to explain factual inconsistencies, leading to clarification questions that may confuse users.
Brevity Trade-offs: Using brevity-inducing prompts reduces token costs significantly (up to 10x) but can strip away the contextual cues necessary for the model to detect ambiguity, slightly impacting accuracy.

Significance and Claims

The paper claims that BAG represents a step toward more robust and honest LLM interactions by allowing models to leverage their own probabilistic uncertainty to guide conversational flow. The authors emphasize that:

Standard LLMs default to answering even when uncertain; BAG corrects this by making uncertainty explicit and actionable.
The ability to abstain based on internal belief states is crucial for factual QA, as many errors stem from a lack of knowledge rather than input ambiguity.
While the method is effective, the distinction between ambiguity and hallucination in the belief state is not yet perfect, and future work should focus on improving the quality of belief states and refining the routing logic.
The approach is a practical, immediate solution for improving QA reliability without the cost of retraining, though it incurs a computational cost of $K$ samples plus the reasoning prompt.

The authors conclude that while BAG improves accuracy and faithfulness, the "disentanglement" of when to clarify versus when to abstain remains an open challenge, particularly given that belief states reflect the model's plausible outputs rather than guaranteed ground truth.