AgentChangeBench: A Multi-Dimensional Evaluation Framework for Goal-Shift Robustness in Conversational AI

AI summary

Overview

Research area: Evaluation methodology for tool-augmented LLM agents in multi-turn conversational settings, specifically robustness to mid-dialogue goal changes.

Technical level: Intermediate. The paper is readable without deep LLM internals, but assumes familiarity with tool-calling agents, multi-turn dialogue evaluation, and metrics such as pass@k.

Scope: AgentChangeBench is a 315-task, three-domain benchmark with five simulated user personas and four complementary metrics that measure whether conversational agents succeed, use tools efficiently, avoid redundant calls, and recover quickly when a user changes goals mid-conversation.

What This Paper Is About

Most benchmarks for LLM agents assume the user's goal stays fixed for the entire conversation, which is unrealistic for real enterprise support. In practice, a customer might start by unlocking a card, then pivot to disputing a fraudulent charge, all in one session. The paper builds a benchmark that deliberately injects these goal shifts and measures not just whether the agent eventually succeeds, but how long it takes to notice the shift, how wasteful it is along the way, and whether its communication quality holds up.

Key Contributions

1. A goal-shift-focused benchmark. AgentChangeBench is presented as the first benchmark explicitly designed to test how LLM agents handle mid-conversation changes in user objectives, rather than static task completion.

2. 315 validated tasks across three enterprise domains. Banking (50), airline (100), and retail (165) tasks, each annotated with an explicit ordered goal sequence, five user personas, and domain-specific APIs. Some templates are reused from τ²-bench, while banking is entirely original.

3. Four complementary evaluation metrics. Task Success Rate (TSR), Tool Use Efficiency (TUE), Tool-Call Redundancy Ratio (TCRR), and Goal-Shift Recovery Time (GSRT), which together capture effectiveness, efficiency, waste, and adaptation latency.

4. A cross-model empirical study. GPT-4o, Claude-3.7-Sonnet, and Gemini-2.5-Flash are evaluated, revealing performance gaps that pass@k scores hide entirely.

Main Findings

• High accuracy does not imply robustness. pass@k frequently drops to 0.0 on new goal-shifted task sets while TSR remains in the 40–60% range, meaning models make substantial partial progress that binary success metrics erase.

• Recovery rates diverge sharply between models. On new airline tasks, GPT-4o recovers from 92.2% of goal shifts, while Gemini-2.5-Flash recovers only 48.6%. Gemini's failure mode is persisting with the prior plan for multiple turns after the user has clearly changed direction.

• Redundancy is severe in retail. Retail-new tasks show a Tool-Call Redundancy Ratio of 89.1% for GPT-4o and 66.5% for Gemini, driven by repeated identical lookups across adjacent turns. Airline redundancy is far lower (13.5–24.1%), so the problem is domain-skewed rather than model-specific.

• Parameter validity is saturated, tool correctness is not. Mean parameter validity P = 0.986 with 98.6% of traces at or above 0.95, while mean tool correctness T = 0.952 with a long tail (4.3% of traces below 0.70). Because P is at ceiling, TUE differences are effectively driven by tool-selection correctness, which motivates reporting TCRR separately.

• Over-confirmation crowds out useful communication. GPT-4o's communication subscore collapses to 11.44% on retail-new tasks despite high action accuracy, as excessive "please confirm" prompts displace information delivery.

• Claude-3.7-Sonnet leads overall, GPT-4o is balanced, Gemini lags. TSR by domain: Claude (banking 57.54%, airline 65.14%, retail 79.57%), GPT-4o (51.25%, 62.19%, 56.48%), Gemini (47.36%, 46.98%, 58.03%). Banking remains the hardest domain for all models.

• Persona difficulty affects outcomes. HARD_1 (suspicious, demanding) yields the lowest TSR at 0.430, while MEDIUM_2 (cost-conscious researcher) reaches 0.580. TUE stays uniformly high across personas (0.946–0.990), meaning personas degrade success and recovery rather than raw tool mechanics. Note: the paper's narrative text and Table 7 report different per-persona recovery figures, so those numbers should be read cautiously.

Methodology in Plain English

The researchers built tasks as declarative JSON specifications. Each task names a persona, lists facts the simulated user knows, and declares an ordered sequence of goals such as ["authentication", "transactions", "dispute"]. The number of required goal shifts equals the number of goals minus one. Shifts are triggered naturally during the conversation—after four user turns on the same goal, after a resolution step, or when the agent asks "anything else?"—and the agent never sees marker text indicating a shift is coming. The simulated user never calls tools; it only discloses known facts, while the agent performs all tool interactions.

Evaluation runs on the τ²-bench harness, which enforces constraints like one tool call per turn and correct goal ordering. Four metrics are computed from the resulting transcripts:

• TSR is a weighted average of three channels: action execution (0.45), behavioral/policy compliance (0.30), and communication quality (0.25). Communication is scored by an LLM judge (gpt-4o-mini) reviewing the full transcript for relevance, clarity, and helpfulness.
• TUE combines tool correctness and parameter validity as 0.6T + 0.4P, with both components reported separately because parameter validity saturates.
• TCRR counts exact duplicate tool calls within a three-turn window, or more than two calls to the same function.
• GSRT measures turns between a user-initiated shift and three subsequent events: agent acknowledgment, first relevant tool call, and evaluator-confirmed goal achievement. A shift counts as recovered only if acknowledgment occurs and no human transfer is attempted.

Tasks were produced through a hybrid pipeline: hand-written exemplars, reused τ²-bench templates, and LLM-generated candidates that were manually reviewed for tool parameter validity, consistency with the domain database, and correct separation of user-known facts from information the agent must surface via tools.

Why This Matters

Impact on research: The paper argues that pass@k collapses distinct behavioral profiles into a single number. An agent that meanders for ten turns and one that adapts immediately can look identical under pass@k but differ enormously under GSRT, TUE, and TCRR. This gives the evaluation community a template for measuring adaptation latency and operational waste separately from raw success.

Real-world applications:

• Enterprise assistant deployment decisions. Organizations can weight metrics to match their priorities—cost-sensitive financial services might optimize TUE and TCRR, while customer experience teams might prioritize GSRT and communication quality.
• Cost control in production agents. Redundancy rates above 80% translate directly into wasted API calls, longer sessions, and higher inference spend.
• Agent training and reward shaping. GSRT and TCRR provide dense, partial-credit signals that binary success rewards do not, useful for fine-tuning or RL.
• Service-level monitoring. Recovery latency after user pivots is a plausible production SLA indicator for conversational agents.

Industry relevance: The domains chosen—banking, retail, and airline—mirror the highest-volume enterprise customer service workflows, and the personas are modeled on realistic user archetypes rather than synthetic extremes. The finding that many tasks produce pass@k = 0.0 with 40–60% TSR suggests that current benchmarks may be systematically underestimating deployed agent usefulness while simultaneously missing real inefficiencies.

Future Directions

• Harder, adversarial personas. The current five personas are relatively benign. The authors propose adding adversarial, deceptive, hostile, or policy-pushing users, conflicting instructions, frequent interruptions, implicit constraints, and multilingual switches.
• Broader tool classes and unified protocols. The benchmark covers only customer-service domain APIs. Extensions would include IDE/code actions, OS/shell control, spreadsheets and BI tools, browsers, and robotics/IoT, ideally through MCP-compatible adapters for a single heterogeneous interface.
• Implicit and ambiguous goal drift. Current shifts are pre-declared and often explicitly signaled. Testing latent drift, overlapping or interleaved objectives, partial reversions, and conflicting goals would probe plan repair under genuine uncertainty.
• Automated task generation and wider model coverage. Scaling to more models (including open-weight ones), more sizes, and additional domains such as healthcare, education, and technical support would improve generality.

Target Audience

Researchers building or evaluating agent benchmarks will find the metric design and the argument against pass@k most valuable. Enterprise AI teams choosing or monitoring conversational agents benefit from the operational framing around redundancy and recovery latency. Conversational AI engineers working on multi-turn planning and tool orchestration can use the failure taxonomy—late shift detection, redundant calls, over-confirmation—as a diagnostic checklist. The paper is most useful to readers already comfortable with tool-calling agent architectures; it is not an introduction to LLM agents.

Authors’ abstract

Goal changes are a defining feature of real world multi-turn interactions, yet current agent benchmarks primarily evaluate static objectives or one-shot tool use. We introduce AgentChangeBench, a benchmark explicitly designed to measure how tool augmented language model agents adapt to mid dialogue goal shifts across three enterprise domains. Our framework formalizes evaluation through four complementary metrics: Task Success Rate (TSR) for effectiveness, Tool Use Efficiency (TUE) for reliability, Tool Call Redundancy Rate (TCRR) for wasted effort, and Goal-Shift Recovery Time (GSRT) for adaptation latency. AgentChangeBench comprises 2,835 task sequences and five user personas, each designed to trigger realistic shift points in ongoing workflows. Using this setup, we evaluate several frontier models and uncover sharp contrasts obscured by traditional $\text{pass}@k$ scores: for example, GPT-4o reaches $92.2\%$ recovery on airline booking shifts while Gemini collapses to $48.6\%$, and retail tasks show near perfect parameter validity yet redundancy rates above $80\%$, revealing major inefficiencies. These findings demonstrate that high raw accuracy does not imply robustness under dynamic goals, and that explicit measurement of recovery time and redundancy is essential. AgentChangeBench establishes a reproducible testbed for diagnosing and improving agent resilience in realistic enterprise settings.