Evaluating LLM-based Agents for Multi-Turn Conversations

Summary: Guan et al. (Microsoft, 2025) survey ~276 papers on evaluation of LLM conversational agents across multi-turn settings. Two interlocking taxonomies: what to evaluate (5 dimensions) and how to evaluate (4 method types). Key finding: traditional single-turn metrics (BLEU, ROUGE) cannot capture the dynamic, contextual nature of multi-turn agent performance.

Sources: Academia/Multi-Turn Conversations.pdf

Last updated: 2026-05-06


Citation

Guan, S., Wang, J., Bian, J., Zhu, B., Lou, J., & Xiong, H. (2025). Evaluating LLM-based agents for multi-turn conversations: A survey. Microsoft Research (China). arXiv:2502.22458.


What is an LLM-based multi-turn conversational agent?

An LLM-based conversational agent adds three capabilities to a base LLM:

  1. Dynamic tool use: autonomously decides when to call external APIs or services (search, booking, databases)
  2. Persistent memory: maintains both turn-level context (within a session) and conversation-level memory (across sessions and user preferences)
  3. Sequential planning: decomposes complex user requests into ordered actionable steps

This is distinct from a plain LLM that responds to isolated prompts. The agent must maintain coherent, goal-directed dialogue across multiple turns — tracking intermediate state, updating plans when new information arrives, and calling the right tools at the right time.

Taxonomy 1 — What to evaluate (5 dimensions)

DimensionCore question
Task completionDid the agent accomplish the user’s goal across all turns?
Response qualityAre individual responses fluent, accurate, and contextually appropriate?
User experienceDoes interaction feel natural, efficient, and satisfying across the session?
Memory and context retentionDoes the agent correctly track facts, preferences, and decisions across turns?
Planning and tool integrationDoes the agent decompose goals correctly and invoke tools appropriately?

Task completion and memory retention are the dimensions where multi-turn agents most distinctively diverge from single-turn LLMs. A hallucination in turn 3 that the agent then builds on in turn 7 produces cascading failure that BLEU scores on individual turns completely miss.

Taxonomy 2 — How to evaluate (4 method types)

MethodDescriptionLimitation
Annotation-basedHuman raters evaluate turns or full conversationsExpensive, slow, inconsistent across raters
Automated metricsBLEU, ROUGE, F1 on reference outputsDesigned for single-turn NLG; miss coherence, goal-tracking, tool-use correctness
HybridCombines human judgment with automated scoringAttempts to get coverage of both; still has referent problem for open-ended agents
Self-judging (LLM-as-judge)A separate LLM evaluates the agent’s outputsScalable, but evaluator-model can share blind spots with the model being evaluated

The LLM-as-judge problem

The survey surfaces a live tension in the field: LLM-as-judge is the most scalable evaluation method, but the evaluating LLM may be structurally unable to catch errors that similarly-trained models make. If both the conversational agent and the judge were trained on similar data and share similar failure modes (e.g., factual hallucinations about specific domains, memory tracking errors), the judge will rate bad outputs as good. This is a variant of the correlated-error problem in test validity — judges must be independent of the system they evaluate.

Historical arc

ELIZA (1960s) → finite-state machines and frame-based systems → hidden Markov models for spoken dialogue → deep neural networks → Transformer architectures → LLM agents. Each generation increased contextual capacity; multi-turn evaluation challenges have scaled accordingly.

Connections

kim-gricean-human-llm addresses the same three-stage interaction cycle (input, process, output) from a pragmatics angle. Guan et al.’s five evaluation dimensions map onto Kim’s maxims: task completion (Quantity + Relation), response quality (Manner), user experience (holistic), memory (Quantity across turns). The two papers are complementary: Kim provides the normative framework (what should happen), Guan provides the measurement infrastructure (how to detect when it doesn’t).

octotools addresses the planning and tool integration dimension specifically — its three-component architecture (tool cards, planner, executor) is a direct engineering response to the evaluation challenges Guan documents.

illusion-of-thinking established that reasoning models collapse at high task complexity. Multi-turn settings compound this: each turn adds to effective complexity, meaning the collapse threshold is reached sooner in a long conversation than in a single-turn problem.