Effective social interaction requires more than fluent responses: agents must reach plans that fit the participants' needs and circumstances. Existing evaluations of social agents mainly rely on LLM judges that rate the conversation, without checking the resulting plan. We introduce HICO-Bench, a benchmark of 180 everyday coordination scenarios, such as resource allocation, scheduling, and authorization, in which two agents with private information converse and then independently record the plan they agreed on. A program checks Feasibility (the two records match and satisfy both participants' private conditions), and an LLM judge checks Integrity (the conversation supports the recorded plan); passing both is a Verified Success. Each scenario has two versions that share the same public context but differ in one private condition, so the correct plan differs between them. Across thirteen LLMs, 36% of conversations rated highly for goal completion fail Verified Success. To learn from outcomes that arrive only at the end of a conversation, we propose Paired-Credit Policy Optimization (PaCPO). After each turn, PaCPO asks the agent for a tentative plan, checks it in both versions of the scenario, and rewards the turns after which the tentative plans fit their respective versions. PaCPO raises Qwen3-8B's Verified Success from 20.4% to 28.3%.
Figures & tables
Figure 1: Reassuring words need not lead to appropriate help. (b) Without clarifying, Leo offers to take over, although Maya can present and only wants a slide review. (c) After clarifying, the appropriate plan depends on whether Maya can attend, which is initially private to her. Feasibility checks that the plan works and both records match; Integrity checks that dialogue supports it.
Interaction family
Example
Private conditions and verification
Resource allocation
Neighbors divide a $240 fence-repair bill.
Shares respect both participants’ bounds and sum to the fixed total.
Authorization
A consultant handles a request for one extra review hour.
Approval stays within delegated authority; otherwise, the request is referred.
Bounded quantities
A buyer and seller agree on the number of apple crates.
The agreed integer quantity lies within both participants’ acceptable intervals.
Joint scheduling
Studio partners arrange a 20-minute celebration.
The shared slot fits both availability windows and has the required duration.
Shared-resource scheduling
Two photographers reserve a tripod for ten minutes each.
Bookings fit both use windows and required durations, without overlapping.
Set-based consent
Two people agree on belief-related discussion topics.
Selected topics match the required set and are all permitted.
Table 1: Interaction families in HICO-Bench. Each family has 16 training, 4 development, and 10 test scenarios, each with two worlds.
Figure 2: PaCPO overview. Rollouts in the two worlds yield a tentative plan after each turn. Checking each plan in both worlds determines turn-level credit, which redistributes the terminal reward before the PPO update. Indices i,j denote worlds, t turns, and k tokens; panel (d) omits i . R is the terminal reward, rt and rt′ are the original and redistributed rewards, and Φt is the potential. The dashed branch carries GAE value targets.
Model
Size
Verifiable outcomes (%)
SOTOPIA ratings
Feasibility
Integrity
Ver. Succ.
Goal
Believ.
Qwen3-8B ( Yang et al., 2025 )
8B
49.58
25.83
20.42
7.47
6.31
MiniMax M3 ( MiniMax, 2026 )
api
58.33
62.50
50.42
9.18
8.68
GPT-4o ( OpenAI, 2024 )
api
62.08
63.33
51.25
9.12
8.71
GPT-5.6 Luna ( OpenAI, 2026a )
api
73.75
56.67
52.08
9.61
8.63
Qwen3.8-27B ( Qwen Team, 2026 )
27B
64.17
62.92
54.58
9.21
8.72
Table 2: Thirteen-model HICO-Bench benchmark. Fixed Qwen3-8B partner, 240 tasks per model; sorted by increasing Verified Success (Ver. Succ.). Size denotes local model size; api denotes remote inference. Goal and Believability range from 0 to 10; full ratings are in Appendix D .
Method
Verifiable outcomes (%)
SOTOPIA ratings
Feasibility
Integrity
Ver. Succ.
Goal
Believ.
Base ( Yang et al., 2025 )
49.58
25.83
20.42
7.47
6.31
PPO ( Schulman et al., 2017 )
50.00
33.75
25.00
7.58
6.51
GRPO ( Shao et al., 2024 )
52.92
31.25
26.25
7.83
6.62
Dr.GRPO ( Liu et al., 2025b )
50.42
27.92
21.25
7.67
6.39
GiGPO ( Feng et al., 2025 )
52.08
30.42
24.58
7.86
6.60
Table 3: Seven-system training comparison. 240 test tasks per row; 384 training dialogues per RL method. To match this budget, group-based baselines ( G=4 ) cover 48 training scenarios, and PPO and PaCPO cover 96. Goal and Believability (0–10) are not optimized. Best values are bold.
Table 4: External interaction tasks. Task-specific training evaluated on 154 MultiWOZ and 86 CaSiNo test dialogues. All metrics are percentages; Norm. Nash denotes normalized Nash surplus.
Figure 5: Scenario-level changes and a dialogue example. (a) PaCPO’s change in successful tests relative to each baseline across 60 scenarios (four tests per column: two worlds and two roles). (b) Verbatim excerpts with the same availability information and Finley as the tested role. Both runs submit the same feasible records, but only PaCPO’s dialogue supports them.
Training condition
Feasibility
Integrity
Ver. Succ.
Same-world candidate control
50.83
32.08
26.25
Shuffled credit positions
50.42
32.92
26.25
Without disclosure gate
49.58
33.33
25.83
PaCPO (full method)
52.92
35.42
28.33
Table 5: PaCPO credit ablations. 240 test tasks per condition; all values are percentages.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Dataset composition. Each family contributes 30 scenarios: 16 training, 4 development, and 10 test. Two worlds per scenario yield 360 worlds; role-wise evaluation yields 96 development and 240 test tasks.
Metric
Range
Mechanism
Interpretation
Feasibility
{0,1}
Rules
Compatible records with terms satisfying the modeled conditions.
Integrity
{0,1}
Language judge
No unresolved factual error, loop, commitment mismatch, or protocol leakage.
Verified Success
{0,1}
Product
Feasibility and Integrity both pass.
Goal
[0,10]
Language judge
Attainment of the tested role’s goal.
Knowledge
[0,10]
Language judge
Acquisition of relevant information.
Believability
[0,10]
Language judge
Naturalness and role consistency.
Appendix
Table 6: Evaluation metrics. Binary outcomes and social dimensions with their scoring mechanisms. The original seven dimensions retain SOTOPIA ranges ( Zhou et al., 2024 ) .
Score
IM: information gaps
GU: interpretation
IA: response appropriateness
0–2
Commits despite a critical manageable unknown, or inquiry seriously blocks progress.
Material fabrication or persistent misunderstanding of facts, roles, or consent.
An unretracted response violates a critical known condition or invents authority.
3–5
Useful exchange occurs, but a consequential gap remains unmanaged.
Only generic acknowledgment, or an unresolved material interpretive error.
A useful direction, but an essential term is missing or unsupported despite an opportunity.
6–8
Main uncertainty handled, with a specific omission, redundancy, or delay.
Main facts understood, with a limited ambiguity or relevant corrected error.
Mainly feasible and concrete, with a limited coordination or scope weakness.
9–10
Demonstrated sufficient information, or a genuinely conditional step managing the dependency.
Concrete responses demonstrate sufficient, accurate understanding and preserve uncertainty.
A sufficiently specified, grounded response; justified refusal or referral can qualify.
Appendix
Table 7: Anchors for the three diagnostic ratings. Each dimension evaluates the tested role over the complete dialogue using the frozen rubric.
Method
Feasibility
Integrity
Ver. Succ.
Base ( Yang et al., 2025 )
49.58
25.83
20.42
SFT
46.25
29.58
25.83
PaCPO
52.92
35.42
28.33
SFT + PaCPO
48.75
36.25
31.67
Appendix
Table 8: SFT initialization and subsequent PaCPO. 240 test tasks per condition. SFT uses 209 accepted teacher dialogues (991 examples, 64 updates); each RL stage uses 384 dialogues. All generated failures remain in the outcome denominator.
Method
SOTOPIA ratings
Diagnostics
Goal
Know.
Believ.
Rel.
Secret
Rules
Mat.
IM
GU
IA
Range
[0,10]
[0,10]
[0,10]
[−5,5]
[−10,0]
[−10,0]
[−5,5]
[0,10]
[0,10]
[0,10]
Base
7.47
7.41
6.31
2.15
-1.43
0.00
1.08
7.87
7.56
7.09
PPO
7.58
7.61
6.51
2.28
-1.24
0.00
1.20
8.12
7.93
7.51
GRPO, G=4
7.83
7.65
6.62
2.24
-1.38
0.00
1.28
8.13
7.89
7.50
Dr.GRPO, G=4
7.67
7.67
6.39
2.25
-1.35
0.00
1.28
8.16
7.97
7.49
Appendix
Table 9: Complete SOTOPIA ratings for training methods. Seven SOTOPIA dimensions and three process diagnostics, reported individually on their original scales. Know., Believ., Rel., and Mat. denote Knowledge, Believability, Relationship, and Material Benefits; Rules denotes Social Rules. All columns are higher-is-better; bold marks column-best values. SFT uses 239 scored dialogues; the other conditions use 240.
Figure 7: Verified outcomes across thirteen models. Each point summarizes 240 tasks; right-hand labels give Verified Success. Complete SOTOPIA ratings appear in Table 10 .
Model
SOTOPIA ratings
Diagnostics
Goal
Know.
Believ.
Rel.
Secret
Rules
Mat.
IM
GU
IA
Range
[0,10]
[0,10]
[0,10]
[−5,5]
[−10,0]
[−10,0]
[−5,5]
[0,10]
[0,10]
[0,10]
Qwen3-8B
7.47
7.41
6.31
2.15
-1.43
0.00
1.08
7.87
7.56
7.09
MiniMax M3
9.18
8.31
8.68
2.80
-0.80
-0.01
1.40
9.31
9.28
9.19
GPT-4o
9.12
8.22
8.71
2.68
-0.76
0.00
1.32
9.06
8.99
8.85
GPT-5.6 Luna
9.61
8.47
8.63
2.38
-0.67
0.00
1.31
9.55
9.62
9.52
Appendix
Table 10: Complete SOTOPIA ratings for thirteen models. Model order follows Table 2 ; abbreviations and score ranges follow Table 9 . Bold marks column-best values. Luna and Sol each have 239 scored dialogues; the other models have 240. All 240 attempts enter the binary-outcome denominators.
Parameter
Value
Policy / fixed partner
Qwen3-8B LoRA / frozen Qwen3-8B
Initialization
Base; no SFT adapter
Train scenarios / worlds / conversations
96 / 192 / 384
Rollout batch / PPO epochs
16 conversations / 2
Actor / value-head learning rate
2×10−6 / 2×10−4
GAE discount / trace
γ=1 / λ=0.95
Appendix
Table 11: Principal PaCPO configuration.
Method
Total output tokens
Probe output tokens
Disclosure requests
Optimizer updates
PPO
276,293
0
0
192
GRPO ( G=4 )
282,431
0
0
188
PaCPO
310,863
57,405
216
192
Appendix
Table 12: Training-generation and annotation cost. Total output includes tested-role, partner, and tentative-plan tokens. Tentative-plan output is a subset of this total; disclosure requests are additional to terminal Integrity scoring.
Figure 8: Why verify outcomes? (a) Verified Success failure rate conditioned on Goal ≥8 , with failed/eligible counts for each model. The dashed line is the pooled rate: 1,010/2,806 model–task observations (35.99%). (b) Base under a 2×2 information-access and dialogue design: 12 development scenarios, 48 tasks per condition. Cells show rule-based Feasibility success; gains are percentage-point differences. Dialogue includes information exchange and negotiation.
Condition
Feasibility
Own-record feasibility
Ver. Succ.
No dialogue
10.42
25.00
–
Revealed + no dialogue
41.67
52.08
–
Hidden + dialogue
60.42
62.50
31.25
Revealed + dialogue
72.92
75.00
39.58
Appendix
Table 14: Information-access and dialogue diagnostic. Conditions in Figure 8 (b); all values are percentages. Own-record feasibility checks the tested role’s record against the full scenario constraints, including both participants’ conditions; Feasibility additionally requires both records to match. Both checks are rule-based. Verified Success here uses the original local Qwen3.8-27B judge, not Qwen3.8-Max.
Method
Feasibility
Integrity
Ver. Succ.
Base ( Yang et al., 2025 )
48.89 ± 2.68
29.17 ± 2.20
21.94 ± 2.10
PPO ( Schulman et al., 2017 )
48.89 ± 2.55
31.67 ± 2.20
27.22 ± 1.73
PaCPO
51.94 ± 2.68
35.83 ± 2.50
28.33 ± 2.20
Appendix
Table 15: Full self-play results. Both participants use the same checkpoint. Mean ± sample SD over three generation seeds; 120 dialogues and 240 scored role views per seed. Rates are percentages; generation failures remain in the denominator.
Figure 9: Results by interaction family. Each cell contains 40 test tasks (ten scenarios, both worlds and roles). Both panels use the same color scale; bold marks the best method within each row and panel. Scheduling remains difficult across methods.
Family
Eligible
Nonzero
Nonzero rate
Resource allocation
272
194
71.3%
Authorization
180
118
65.6%
Bounded quantities
264
194
73.5%
Joint scheduling
222
95
42.8%
Shared-resource scheduling
244
85
34.8%
Set-based consent
196
80
40.8%
Appendix
Table 16: Paired-credit activation. Eligible positions have an open gate and two valid tentative plans; nonzero positions satisfy Mii=Mi,1−i . Training uses 384 dialogues over 24 rounds.
Judge pair
Agreement (%)
Cohen’s κ
Qwen3.8-Max pass (%)
Kimi-K3 pass (%)
Qwen3.8-Max / Kimi-K3
90.00
0.783
30.00
40.00
Appendix
Table 17: Qwen3.8-Max–Kimi-K3 Integrity agreement. Both judges evaluate the same 240 model–task cases. Agreement compares final pass/fail decisions; pass rates show each judge’s marginal acceptance rate.
Evaluating LLM-powered interactive social agents is challenging because socially relevant behaviors depend not only on isolated outputs, but also on prior interactions, social roles, and downstream actions. Existing methods typically allow a target agent to act freely in an environment and then score the resulting trajectory. However, this passive setup can miss capabilities that only become observable under specific social circumstances; for example, conflict handling may remain untested if no disagreement arises. We propose Online Agent-as-a-Judge, a situation-generating evaluation framework for interactive social agents. Online Agent-as-a-Judge deploys an in-world evaluator agent that interacts with the target agent through the environment's native dialogue and action protocol, actively eliciting situations relevant to the evaluation criteria. The resulting trajectories provide evidence for assessing both immediate responses and subsequent behavior. In a life-simulation environment with 32 designer-authored social criteria, Online Agent-as-a-Judge improves criteria coverage and agreement with human labels, yielding more reliable evidence-grounded evaluations of behaviors that passive methods can leave unobserved.
Large language model (LLM) agents are increasingly deployed as teams, yet they fail in characteristically social ways: information holders stay silent, dominant speakers anchor groups on wrong answers, and norm violations go unsanctioned. We present Agent Sociodrama, a framework for modeling and validating evolving social dynamics in LLM agent teams. Under fixed personas, agents' emotional states and directed, asymmetric relationships are updated every round by a unified external LLM judge and shape their subsequent interactions. Across six LLM backbones and five social scenarios, we assess social fidelity through classical social-theory patterns and agreement with human judgments. The resulting trajectories exhibit structured social patterns and model-specific differences. We then introduce Trajectory-Guided Persona Recasting (TPR), which diagnoses group-dynamic failures from interaction transcripts and social trajectories and recasts agent personas before rerunning the task. Across four collaborative benchmarks, TPR improves performance from both heterogeneous and randomly initialized agent persona compositions, yielding up to 83.3%p of accuracy improvement over the initial casting. On representative distributed-information cooperation and mixed-motive bargaining tasks, we outperform SOTOPIA and AgentVerse in three of four comparisons, by up to 37.5%p in Pareto rate. These results show that validated social trajectories provide actionable signals for repairing LLM agent teams.
Yirui Qi, Xiaoming Zhang, Jaehong Yoon
Beihang University · Nanyang Technological University, Singapore
With the rise of personalized, persistent LLM agent frameworks such as OpenClaw, human-centered agentic social networks in which teams of collaborative AI agents serve individual users in a social network across multiple domains are becoming a reality. This setting creates novel privacy challenges: agents must coordinate across domain boundaries, mediate between humans, and interact with other users' agents, all while protecting sensitive personal information. While prior work has evaluated multi-agent coordination and privacy preservation, the dynamics and privacy risks of human-centered agentic social networks remain unexplored. To this end, we introduce AgentSocialBench, the first benchmark to systematically evaluate privacy risk in this setting, comprising scenarios across seven categories spanning dyadic and multi-party interactions, grounded in realistic user profiles with hierarchical sensitivity labels and directed social graphs. Our experiments reveal that privacy in agentic social networks is fundamentally harder than in single-agent settings: (1) cross-domain and cross-user coordination creates persistent leakage pressure even when agents are explicitly instructed to protect information, (2) privacy instructions that teach agents how to abstract sensitive information paradoxically cause them to discuss it more (we call it abstraction paradox). These findings underscore that current LLM agents lack robust mechanisms for privacy preservation in human-centered agentic social networks, and that new approaches beyond prompt engineering are needed to make agent-mediated social coordination safe for real-world deployment.