🤖 AI

Uncertainty Decomposition for Clarification Seeking in LLM Agents

This paper addresses the limitations of classical uncertainty frameworks for interactive LLM agents by proposing a prompt-based decomposition that separates action confidence from request uncertainty to enable proactive clarification seeking, which significantly outperforms existing methods on benchmarks featuring deliberately underspecified tasks across multiple model backbones.

Original authors: Gregory Matsnev

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gregory Matsnev

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

Imagine you are hiring a very smart, but slightly overconfident, personal assistant to run errands for you. You give them a list of instructions, like "Buy a shirt" or "Put the mug on the desk."

Sometimes, you forget to tell them important details, like the color of the shirt or which desk. In the past, if this assistant didn't know what to do, they would either:

Guess wildly and buy a red shirt when you wanted blue, then confidently tell you, "I'm 90% sure this is right!" (This is a mistake).
Say they are confused but give no reason why, making it hard for you to know if they are just struggling with a hard task or if you actually forgot to give them enough info.

This paper introduces a new way to train these AI assistants to handle that confusion better. Here is the breakdown using simple analogies:

The Core Problem: One Number Isn't Enough

Currently, most AI assistants give you a single "confidence score" (like a grade from 0 to 100) for every action they take.

The Flaw: If the assistant says, "I'm only 40% confident," you don't know why.
- Is it because the task is hard? (e.g., "There are 500 shirts to choose from, and I'm not sure which one you like.")
- Or is it because you didn't give enough info? (e.g., "You told me to buy a shirt, but you didn't say what color or size!")

If the assistant is confused because you forgot to give details, they shouldn't just guess; they should ask you, "Hey, what color do you want?" The old system couldn't tell the difference between "This is hard" and "You forgot to tell me something."

The Solution: Splitting the "Confidence" in Two

The authors propose a simple trick: instead of asking the AI for one confidence score, they ask for two separate scores in its internal thinking process:

Action Confidence: "How sure am I that this specific action will work, assuming I understand the goal correctly?"
- Analogy: "I'm 90% sure I can drive to the store, but I don't know which store you want."
Request Uncertainty: "How much information is missing from your instructions?"
- Analogy: "You didn't tell me the store name or the color. I'm 100% unsure because you left out the details."

By separating these two, the AI gets a dedicated "channel" to say, "I know what to do, but I need more info from you," rather than just saying, "I'm confused."

The Experiment: The "Missing Details" Test

To prove this works, the researchers created two new video-game-like tests (WebShop and ALFWorld) where they deliberately hid 50% of the instructions.

The Setup: They told the AI, "Buy a shirt," but removed the words "black" and "size large" from the prompt.
The Goal: They wanted to see if the AI would stop and ask, "What color and size?" instead of just guessing.

They tested this new "Two-Score" method against older methods (which only used one score) using five different powerful AI models.

The Results

Better at Asking: The new method was much better at spotting missing information. On average, it improved the ability to ask for clarification by 73% compared to the old methods. It successfully identified missing details in almost every AI model tested.
No Loss in Skill: Crucially, giving the AI this extra "thinking step" didn't make it worse at actually doing the tasks when the instructions were clear. It didn't lose its ability to solve problems; it just got better at knowing when to stop and ask.

The Catch: The "Brain Drain"

The paper also found a downside, which they call "Capability Dilution."

The Metaphor: Imagine the AI has a fixed amount of mental energy (like a battery). Usually, it uses 100% of that battery to solve the task. Now, it has to split that battery: some for solving the task, some for checking its confidence, and some for checking if you gave enough info.
The Result: Because it's splitting its energy, it sometimes does the actual task slightly worse than before. It's like a runner who is great at sprinting but slows down a tiny bit because they are also trying to check their watch and talk to the coach at the same time.

The Big Picture Conclusion

The authors conclude that while this "Two-Score" trick works well right now, it's a bit of a hack. It forces the AI to do extra thinking just by changing the words in the prompt.

They argue that the future isn't about better prompts, but about retraining the AI itself. They believe the next generation of AI should be built from the ground up to naturally understand the difference between "This is hard" and "You forgot to tell me something," without needing to be tricked by a prompt.

In short: The paper shows that if you teach an AI to separate "I don't know the answer" from "You didn't give me the question," it becomes much better at asking for help. But doing this via simple text instructions costs a little bit of its overall brainpower.

Technical Summary: Uncertainty Decomposition for Clarification Seeking in LLM Agents

1. Problem Statement

Interactive Large Language Model (LLM) agents operate under partial observability, receiving underspecified natural-language instructions and navigating noisy environments. Small errors in early reasoning steps can propagate, leading to confidently incorrect outcomes. While uncertainty estimation is a standard mitigation tool, existing frameworks (specifically the aleatoric/epistemic dichotomy) are argued to be insufficient for interactive agents. Kirchhof et al. [13] and Kim et al. [14] posit that agents require underspecification-aware uncertainty representations that are decomposed by source and communicable to users to enable proactive capabilities like clarification seeking.

Furthermore, practical deployment constraints—such as black-box API access, strict latency budgets, and the absence of labeled training trajectories—render logprob-based, multi-sampling, and training-based uncertainty methods infeasible. This leaves prompt-based methods (where the agent emits uncertainty as structured text) as the only viable approach for real-world deployment. However, existing prompt-based methods typically output a single scalar confidence score, conflating two distinct sources of uncertainty:

Task Difficulty: The action is hard to execute (e.g., choosing from many similar products).
Goal Ambiguity: The user's request is underspecified (e.g., "find a shirt" without color/size).

Conflating these signals prevents the agent from distinguishing between "proceed cautiously" and "ask for clarification."

2. Methodology

Proposed Decomposition

The paper proposes a prompt-based decomposition that separates the single confidence scalar into two distinct signals emitted at each step $t$ :

Action Confidence ( $c_t$ ): The agent's confidence that the chosen action moves toward task completion, given the current understanding. This mirrors the scalar confidence used in baselines.
Request Uncertainty ( $u_t$ ): An estimate of whether the user's goal is fully specified, ranging from 0 (fully specified) to 1 (critical details missing). This is anchored on a three-point scale:
- $0$: Exactly one correct interpretation.
- $0.5$: At least one parameter is open/unstated.
- $1$: Critical details missing; many valid interpretations exist.

Clarification Trigger

The method introduces a deterministic routing mechanism. If $u_t \geq \theta$ (where $\theta = 0.5$ by default), the agent is instructed to emit the special action request_clarification instead of a task action. This creates a direct link between the uncertainty estimate and an observable behavior, allowing the system to be evaluated as a binary classifier for underspecification.

History Propagation

Following the Uncertainty-Aware Memory (UAM) framework [21], both the uncertainty values ( $u_t, c_t$ ) and their natural language explanations are appended to the interaction history $H_t$ . This allows subsequent steps to reason about accumulated request-level uncertainty and action confidence.

Baselines

The proposed method is compared against two prompt-based baselines:

ReAct+UE: A standard ReAct agent with a prompt suffix asking for a single confidence score and explanation. It does not retain uncertainty in the history.
Uncertainty-Aware Memory (UAM): Similar to ReAct+UE but propagates the confidence score and explanation through the history, allowing the agent to reason about past uncertainty.

Benchmarks

To evaluate clarification seeking, the authors constructed two clarification-augmented benchmarks where 50% of tasks are deliberately underspecified:

WebShop-Clarification: Attributes (e.g., color, material) and option clauses are stripped from instructions.
ALFWorld-Clarification: Objects or receptacles are randomly removed from the goal (e.g., "put something in container").
In these variants, if the agent correctly requests clarification on an underspecified task, the original goal is revealed, and the episode continues, ensuring fair success rate measurement.

Evaluation also includes standard benchmarks (WebShop, ALFWorld, REAL) for fault detection (predicting trajectory failure) and task success rate.

3. Key Contributions

Comparative Analysis: A systematic study demonstrating that practical constraints leave prompt-based methods as the only viable family for uncertainty in black-box LLM agents, motivating a deeper study of their capabilities for proactive behaviors.
New Benchmarks: The introduction of WebShop-Clarification and ALFWorld-Clarification, which enable the evaluation of clarification seeking as a binary classification task.
Decomposition Method: A prompt-only method that separates action confidence from request uncertainty, enabling proactive clarification without training or model access.
Empirical Evaluation: A comprehensive evaluation across five LLM backbones (GPT-5.1, DeepSeek-v3.2-exp, GLM-4.7, Qwen3.5-35B, GPT-OSS-120B) on both clarification-seeking and fault-detection tasks.

4. Results

Clarification Seeking

The proposed decomposition significantly outperforms baselines in identifying underspecified tasks:

ALFWorld-Clarification: The proposed method improves clarification F1 by 73% over ReAct+UE and 36% over UAM when averaged across the five backbones.
WebShop-Clarification: The proposed method leads in clarification F1 on every backbone.
Generalization: The gains generalize across different model families, indicating the decomposition is not specific to a single LLM architecture.

Fault Detection

The decomposition does not sacrifice the ability to detect faults (predicting trajectory failure):

Across all backbones and benchmarks, the proposed method achieves fault-detection ROC-AUC scores comparable to ReAct+UE and UAM.
On WebShop and REAL, the proposed method achieves the highest ROC-AUC for last-step and average aggregations.

Task Success Rate and Calibration

Capability Dilution: A monotonic decline in task success rate was observed as uncertainty instrumentation increased: ReAct+UE (28.6%) > UAM (27.8%) > Proposed (27.0%). The authors attribute this to "capability dilution," where the fixed reasoning budget is split between task solving and multiple uncertainty assessments.
Overconfidence: Reliability diagrams show systematic overconfidence across all methods; predicted confidence consistently exceeds observed success rates (ECE ranges from 0.24 to 0.66).
Aggregation Sensitivity: The choice of trajectory-level aggregation (e.g., product, average, last-step) significantly impacts fault detection scores. Notably, on ALFWorld, product aggregation acts as a proxy for trajectory length rather than a true confidence signal, as failed trajectories are systematically longer.

5. Significance and Claims

The paper claims that the proposed decomposition successfully operationalizes the concept of "underspecification uncertainty" as a distinct category, addressing calls from Kirchhof et al. [13] and Kim et al. [14]. By providing a dedicated channel for goal ambiguity, the agent can proactively seek clarification rather than silently failing or proceeding with arbitrary interpretations.

However, the authors are modest about the limitations of the prompt-based paradigm:

Prompt-Based Limits: The observed capability dilution and systematic overconfidence are inherent to the prompt-based approach, where the model must simultaneously solve the task and assess its own confidence within a fixed token budget.
Future Direction: The paper argues that the most promising path forward is native uncertainty estimation. Rather than prompting for decomposed scores, models should be trained (via RLHF, DPO, or GRPO) to internally distinguish between task difficulty and goal ambiguity. This would eliminate the need for prompt-based elicitation, remove capability dilution, and allow for calibrated, natural-language communication of uncertainty.

In conclusion, the work serves as a proof of concept that decomposing uncertainty enables new agent capabilities, but it posits that the next step requires moving these mechanisms from the prompt into the model itself through alignment-based training.