Reasoning Gets Harder for LLMs Inside A Dialogue

1. Problem Statement

While Large Language Models (LLMs) demonstrate strong performance on isolated reasoning benchmarks (e.g., math, logic puzzles), their ability to reason effectively within Task-Oriented Dialogue (TOD) settings remains under-explored. In real-world TOD scenarios, models must simultaneously:

• Perform complex reasoning (arithmetic, spatial, temporal).
• Adhere to strict constraints regarding role, tone, format, and style.
• Manage multi-turn interactions and tool usage (function calling).

The authors hypothesize that the "dialogue framing" itself introduces a performance penalty, causing models to underperform compared to their isolated task capabilities. Existing benchmarks often rely on multiple-choice questions or lack the dynamic, multi-turn nature of real interactions, failing to capture these specific challenges.

2. Methodology: The BOULDER Benchmark

To investigate this gap, the authors introduced BOULDER (Benchmarking of Usefulness of LLMs in Dialogue-Embedded Reasoning).

• Dynamic & Contamination-Resistant: The benchmark is dynamically generated. Examples are synthesized on-the-fly using templates and a database (adapted from MultiWOZ), ensuring that specific instances are unlikely to exist in training data.
• Dual-Setting Design: Each problem instance is presented in two variants:
  1. Isolated (Baseline): A single prompt containing the data and the question.
  2. Dialogue (TOD): The same problem embedded in a multi-turn conversation history, including previous user/assistant turns, tool calls, and retrieved JSON data.
• Task Scope: The benchmark covers 8 distinct tasks across four domains (trains, hotels, restaurants, attractions), testing:
  • Arithmetic: Ticket/booking price calculations with complex constraints.
  • Temporal: Departure times relative to sunset, frequency calculations, and opening hour overlaps (using Allen's interval algebra).
  • Spatial: Distance calculations (Euclidean), directional relations (N/S/E/W), and shortest path planning (Traveling Salesman Problem variant).
• Evaluation Pipeline:
  • LLM-based Parsers: Since models generate natural language, specialized LLMs (Qwen3 30B) are used to extract structured answers (JSON) from the generated text.
  • Bias Correction: To account for parser errors, the authors apply Prediction-Powered Inference (PPI), using human-annotated validation sets to correct point estimates and confidence intervals.
  • Metrics: Accuracy, Precision, and Mean Absolute Error (MAE), normalized to a 0–1 scale.

3. Experimental Setup

• Models Evaluated: 8 LLMs of varying sizes and architectures (4B to 671B parameters), including both open-weight (Qwen3, Mistral, Command A, DeepSeek) and proprietary models (Gemini 2.5 Flash, Claude 4.5 Sonnet).
• Settings:
  1. Baseline: Isolated task solving.
  2. Dialogue: Standard multi-turn TOD with tool usage.
  3. Dialogue-Concise: Dialogue with strict length constraints (max 2 sentences).
• Ablation Studies: To isolate the causes of performance degradation, the authors conducted ablations on:
  • Task Complexity: Removing unused domains/tools.
  • Tool Usage: Removing tool schemas and moving data to the prompt.
  • Turn Structure: Converting multi-turn dialogue to single-turn and vice versa.
  • Role Conditioning: Adding assistant persona instructions to the baseline.
  • Reasoning Instructions: Explicitly prompting models to "take more time" for reasoning.

4. Key Results

• Significant Performance Gap: There is a substantial and consistent drop in performance when moving from the Baseline to the Dialogue setting across all models.
  • Example: Qwen3 4B dropped from ~0.83 (Baseline) to ~0.20 (Dialogue) on the "Ticket Price" task. Even top-tier models like Claude 4.5 Sonnet showed significant degradation.
  • The gap is often more pronounced in smaller models but remains statistically significant for large models.
• Impact of Multi-Turn Interaction: The Multi-turn Baseline ablation revealed that simply adding conversation history (even without role instructions) significantly lowers scores. Conversely, converting the Dialogue setting to a Single-turn variant improved performance, confirming that iterative generation is a primary bottleneck.
• Role Conditioning: Adding "assistant persona" instructions (e.g., "be friendly and helpful") to the isolated baseline caused a performance drop similar to the full dialogue setting, suggesting that persona bias forces models to prioritize style over reasoning.
• Tool Usage: Removing tools from the dialogue context (moving data directly to the prompt) improved scores, though not to baseline levels. This suggests the dual burden of generating text and managing tool schemas degrades reasoning.
• Response Length: Models in dialogue settings generated significantly shorter responses than in the baseline, often skipping Chain-of-Thought (CoT) reasoning and providing "early" answers with weak justifications.
• Failure Modes: Qualitative analysis showed models in dialogue settings frequently:
  • Refuse to solve tasks claiming lack of tools.
  • Misinterpret coordinate systems (spatial tasks).
  • Provide suboptimal solutions (e.g., picking the first valid train instead of the latest).

5. Key Contributions

1. BOULDER Benchmark: A novel, dynamic benchmark specifically designed to evaluate reasoning within realistic, multi-turn TOD scenarios, addressing data contamination and the lack of interactive evaluation in current benchmarks.
2. Empirical Evidence of the "Dialogue Penalty": The study provides robust evidence that framing reasoning tasks within a dialogue context significantly degrades LLM performance, a phenomenon not captured by standard isolated benchmarks.
3. Diagnosis of Failure Modes: Through ablation studies, the authors identify multi-turn interaction, role conditioning, and tool-use requirements as the primary drivers of this performance gap.
4. Methodological Rigor: The use of PPI for bias correction in LLM-based parsing sets a standard for reliable evaluation of natural language outputs.

6. Significance and Implications

• Re-evaluating Benchmarks: The results suggest that high scores on isolated reasoning benchmarks (like MATH or GSM8K) may overestimate an LLM's robustness in real-world applications where reasoning is embedded in conversation.
• System Design: For developers building TOD agents, relying on a single LLM to handle both reasoning and dialogue management (zero-shot) may be insufficient. The findings highlight the need for specialized architectures or evaluation protocols that account for the "reasoning cost" of maintaining a persona and managing context.
• Future Directions: The paper calls for evaluating LLMs in realistic interactive scenarios rather than static prompts. It also suggests that current "reasoning models" (which separate thought from output) might struggle in TOD due to latency and cost constraints, necessitating new approaches to integrate reasoning into conversational flows without sacrificing performance.

In conclusion, the paper demonstrates that context matters: the very structure of a dialogue (multi-turn, role-playing, tool integration) acts as a cognitive load that hinders the logical capabilities of even state-of-the-art LLMs.