cs.AIAug 28, 2026
SaveAcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics
Organizations: University of Southern California · Contextual AI · Capital One · Georgia Tech University
Abstract
User preferences in user-agent collaboration are rarely static and fully-specified upfront: preferences are formed, revealed, adjusted, and relaxed during interaction. Existing benchmarks for evaluating user-agent collaboration focus almost exclusively on resolving underspecified preferences, thereby failing to capture the richer dynamics of real-world interaction. We introduce AcCoRD, a user-agent collaboration benchmark requiring agents to handle diverse user preference dynamics in two domains: online shopping and travel planning. We evaluate five frontier LLMs under two prompting strategies: vanilla ReAct, and an uncertainty-guided variant that prompts models to identify and resolve ambiguity about user preferences. Our results reveal that frontier models can handle underspecification but struggle to satisfy preferences that emerge or evolve mid-interaction and require more sophisticated uncertainty modeling. Further, prompting alone fails to elicit the required uncertainty recognition. We release AcCoRD as a resource for developing agents that can navigate the full complexity of real-world user preferences.
Figures & tables
Figure 1: In AcCoRD agents must elicit and navigate user preferences that may be withheld, flexible, unachievable, or not yet formed at the start of the interaction.
| Benchmark | Domain(s) | Hard Underspec. | Soft Underspec. | Unachievable | Triggered |
| ChatShop [ 4 ] | Shopping | ✓ | ✗ | ✗ | ✗ |
| -bench [ 34 ] | Customer service | ✓ | ✗ | ◐ | ✗ |
| ToolSandbox [ 14 ] | Mobile use | ✓ | ✗ | ✗ | ✗ |
| Ambig-SWE [ 27 ] | Coding | ✓ | ✗ | ✗ | ✗ |
| UserBench [ 18 ] | Travel booking | ✓ | ◐ | ✗ | ✗ |
| ColBench [ 39 ] | Software engineering | ✓ | ✗ | ✗ | ✗ |
Table 1: Comparison of benchmarks requiring user-agent interaction across four preference dynamics. ✓ = dynamic fully represented in benchmark, ◐ = partially represented, ✗ = not represented.
Figure 2: Sample tasks from both AcCoRD benchmarks, demonstrating how the different preference benchmarks are represented in the tasks and presented to the user simulator.
| Preference Dynamics (avg / %) | ||||||
| Domain | Search Space | Avg. Prefs per Scenario | Hard Underspec. | Soft Underspec. | Unachievable | Triggered |
| Shopping | 1.2M products | 0 4.89 | 2.80 (57.3%) | 0.75 (15.3%) | 0.99 (20.2%) | 0.35 (7.2%) |
| Travel | 14 options/domain | 11.25 | 4.99 (44.4%) | 2.23 (19.8%) | 2.28 (20.3%) | 1.75 (15.6%) |
Table 2: Dataset statistics for the AcCoRD test sets (100 scenarios each). For each preference dynamic, we report the average count per scenario and its percentage of all preferences.
| Model | Method | Task Completion Rate (%) | Outcome Score | Perfect Outcome Rate (%) | Task Score | |
| AcCoRD -Shop | Llama3.1-70B | ReAct | 72.0% | 0.644 | 17.0% | 0.464 |
| UncReAct | 56.0% | 0.703 | 15.0% | 0.394 | ||
| DeepSeek-V3.2 | ReAct | 93.0% | 0.688 | 20.0% | 0.640 | |
| UncReAct | 83.0% | 0.729 | 20.0% | 0.605 | ||
| GPT-4.1 | ReAct | 91.3% | 0.675 | 15.7% | 0.619 | |
| UncReAct | 94.7% | 0.663 | 19.0% | 0.629 |
Table 3: Main results on AcCoRD , averaged across 100 scenarios with 3 rollouts per scenario.
| Dynamic-wise Preference Score | Rollout stats (mean/median) | ||||||||
| Model | Method | Hard Underspec. | Soft Underspec. | Unachievable | Triggered | # env. actions | # utterances | ||
| AcCoRD -Shop | Llama3.1-70B | ReAct | 0.739 | 0.921 | 0.614 | 0.091 | 17.79 / 14 | 4.56 / 4 | |
| UncReAct | 0.796 | 0.92 | 0.704 | 0.222 | 4.70 / 4 | 10.21 / 9 | |||
| DeepSeek-V3.2 | ReAct | 0.709 | 0.871 | 0.674 | 0.147 | 17.67 / 15 | 1.29 / 0 | ||
| UncReAct | 0.777 | 0.905 | 0.73 | 0.175 | 16.07 / 12 | 4.05 / 3 | |||
| GPT-4.1 | ReAct | 0.737 | 0.878 | 0.679 | 0.186 | 14.39 / 9 | 1.84 / 1 | ||
Table 4: Preference scores for each preference dynamic type on AcCoRD -Shop and AcCoRD -Travel, averaged across all trajectories where the agent completed the task. We also report the mean and median number of environment actions taken and utterances made by the agent in each setting.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Test | AcCoRD -Shop prompt | AcCoRD -Travel prompt |
| test000: Response relevance | Does the candidate response engage with at least one thing the assistant said or asked in its last message? ( yes if the response addresses something from the assistant’s last message; no if the response is about an unrelated topic) | Does the candidate response engage with at least one thing the assistant said or asked in its last message? ( yes if the response addresses something from the assistant’s last message; no if the response is about an unrelated topic) |
| test001: Preference value consistency | If the assistant asks the user for a preference value and the user’s instructions specify a value for that preference, are all preference values stated in the candidate response consistent with the user’s instructions ? ( yes if every stated value matches the instructions; no if any stated value contradicts the instructions; n/a if the assistant did not ask for a preference value, or the response did not address the request) | If the assistant asks the user for a preference value and the scenario specifies a value for that preference, are all preference values stated in the candidate response consistent with the scenario ? ( yes if every stated value matches the scenario; no if any stated value contradicts the scenario; n/a if the assistant did not ask for a preference value, or the response did not address the request) |
| test002: Contradiction acknowledgment | If the assistant describes a specific product whose attribute contradicts a preference the user has already revealed earlier in this conversation, does the candidate response point out that contradiction? ( yes if the contradiction is pointed out; no if the contradiction is ignored or the response accepts the product; n/a if the assistant did not describe a specific product, or no described attribute contradicts an already-revealed preference) | If the assistant describes a specific hotel/flight/activity option whose attribute contradicts a preference the user has already revealed earlier in this conversation, does the candidate response point out that contradiction? ( yes if the contradiction is pointed out; no if the contradiction is ignored or the response accepts the option; n/a if the assistant did not describe a specific option, or no described attribute contradicts an already-revealed preference) |
| test003: Backoff acceptance | If the assistant has indicated that the user’s primary preference is unavailable AND the user’s instructions define a backoff for that preference, does the candidate response accept the backoff rather than continuing to insist on the unavailable primary? ( yes if the response accepts the backoff or otherwise shows appropriate flexibility; no if the response continues to insist on the unavailable primary; n/a if the assistant has not indicated that the primary is unavailable, or the user’s instructions define no backoff for that preference) | If the assistant has indicated that the user’s primary preference is unavailable AND the scenario defines a backoff for that preference, does the candidate response accept the backoff rather than continuing to insist on the unavailable primary? ( yes if the response accepts the backoff or otherwise shows appropriate flexibility; no if the response continues to insist on the unavailable primary; n/a if the assistant has not indicated that the primary is unavailable, or the scenario defines no backoff for that preference) |
| test004: Premature backoff disclosure | If the candidate response reveals a backoff preference, had the assistant already indicated earlier in the conversation that the primary preference could not be satisfied? ( yes if the backoff is only revealed after the primary was confirmed unavailable; no if the backoff was revealed before the primary was confirmed unavailable; n/a if the response does not reveal a backoff preference) | If the candidate response reveals a backoff preference, had the assistant already indicated earlier in the conversation that the primary preference could not be satisfied? ( yes if the backoff is only revealed after the primary was confirmed unavailable; no if the backoff was revealed before the primary was confirmed unavailable; n/a if the response does not reveal a backoff preference) |
| test005: Triggered preference revelation | If the user’s instructions include a preference marked ‘only reveal if mentioned’ AND the assistant’s last message explicitly surfaces the specific target value for that preference (e.g., names the exact value, or presents a product whose attribute equals that value), does the candidate response reveal that preference? ( yes if the matching preference is revealed; no if the response withholds or misstates the preference when it should have been revealed; n/a if no such preference applies, or the assistant did not surface the specific target value) | If the scenario includes a preference marked ‘only reveal if mentioned’ AND the assistant’s last message explicitly surfaces the specific target value for that preference (e.g., names the exact value, or describes an option whose attribute equals that value), does the candidate response reveal that preference? ( yes if the matching preference is revealed; no if the response withholds or misstates the preference when it should have been revealed; n/a if no such preference applies, or the assistant did not surface the specific target value) |
Table 5: Unit test prompts for the AcCoRD -Shop and AcCoRD -Travel user simulators. Differences between benchmarks are bolded . All tests return yes , no , or n/a .
| AcCoRD -Shop | AcCoRD -Travel | ||||
| Claude-4.5 | GPT-5.1 | Claude-4.5 | GPT-5.1 | ||
| test000: Response relevance check | 100.0% | 100.0% | 98.1% | 94.7% | |
| test001: Preference value consistency | 100.0% | 98.9% | 100.0% | 91.0% | |
| test002: Contradiction acknowledgment | 93.4% | 96.8% | 56.1% | 77.7% | |
| test003: Backoff acceptance | 100.0% | 96.8% | 90.1% | 88.9% | |
| test004: Premature backoff disclosure | 100.0% | 83.6% | 97.9% | 87.5% | |
Table 6: Unit test pass rates for two candidate LLMs, Claude-Sonnet-4.5 and GPT-5.1, on both benchmarks. Table 5 contains prompts for evaluating the different unit tests.
| You are a helpful assistant that evaluates the quality of a product purchase. |
| You will be given a product that has been purchased by a shopping assistant, and a criteria for the product purchase. You need to provide a binary score (0 or 1) based on whether the purchased product meets the criteria. If the criteria is very specific (e.g. a t-shirt that is a shade/pattern of black specific to a product), and the product is similar enough (e.g. a t-shirt that is also a shade/pattern of black), you should give a score of 1. |
| The purchased product: Product name: {product_name} Product description: {product_description} Product price: {product_price} Product attributes: {product_attributes} Product features description: {product_features_description} Selected options when purchasing: {selected_options} |
| The criteria for the product purchase: {criteria} |
Table 7: Prompt provided to the Claude-Sonnet-4.5 for evaluating
More in this topic
LLM Agent Evaluation
Large language model agent evaluation measures whether tool-using systems complete tasks reliably, safely, and efficiently.
See all 1,132 papers