Delay-of-Gratification as a Multi-Agent Survival Micro-benchmark for Long-Horizon LLMs: Social Exposure, Personas, and Tool Use Budgets

Olga Manakina; Igor Bogdanov; Chung-Horng Lung

cs.AIAug 26, 2026

Delay-of-Gratification as a Multi-Agent Survival Micro-benchmark for Long-Horizon LLMs: Social Exposure, Personas, and Tool Use Budgets

Authors: Olga Manakina, Igor Bogdanov, Chung-Horng Lung

Organizations: Department of Cognitive Science Carleton University, Ottawa, ON K1S 5B6 · Systems & Computer Engineering Carleton University, Ottawa, ON K1S 5B6

Abstract

Large language models (LLMs) are increasingly deployed as multi-turn agents that must sustain goals, use tools, and adapt to other agents over extended interactions. However, existing research lacks auditable, multi-turn, multi-factorial experiments that quantify LLM behavior under explicit constraints, with time-resolved statistics that reveal how behavior unfolds over long horizons. To address this gap, we develop a multi-agent micro-benchmark inspired by the Stanford marshmallow experiment: ReAct agents operate minute-by-minute with a "raise a question" tool under a per-step budget, while we factorially manipulate social context (broadcast vs. isolated), personas (age, hedonic drive), and metacognitive policy (mandatory vs. optional tool use). We analyze outcomes with Kaplan-Meier (KM) survival curves and discrete-time hazard models over a long risk horizon across 19,200 agent trajectories in 64 cells. Behavior shows a sharp early "eat" impulse, and only 75.9% of agents persist to the end. In a discrete-time hazard model, isolation reduces per-minute risk relative to broadcast, whereas a must-use self-questioning policy increases risk. On average, agents ask $\approx 7.12$ questions and hit the per-step budget in $\approx 6\%$ of minutes. Questioning declines faster under broadcast than isolation. Ablation experiments demonstrated that removing hedonic drive and/or persona age increases survival and completion, narrows the broadcast/isolated gap, but leaves the must vs. may ordering intact. The combined ablation (no hedonic + no persona age) yields the highest completion (approaching $1.0$ ). These results establish delay-of-gratification as a compact, multi-turn interaction benchmark that captures social contagion and tool-use dynamics in LLM agents, providing a reproducible testbed and statistics for analyzing long-horizon, multi-agent behavior.

Figures & tables

Panel of Figure 1 — Figure 1: Interaction loops for (a) the partially observable Markov decision process (POMDP, broadcast condition) and (b) the fully observable Markov decision process (MDP, isolated condition). In (a), the agent observes time and peer responses, introducing partial observability. In (b), the agent observes only time and its own history, rendering the environment fully observable. At each step, the agent internally uses a capped raise_a_question tool, then chooses an action $a_{t}\in\{\text{wait},\text{eat}\}$ . The environment provides a reward, updates the state, and advances to the next observation.

Data Summary		Survival Metrics (Overall)		Data Quality
Agents per Cell	6	Initial Eat Rate	0.125	Data Quality Rate	99.9%
Total Agent Trajectories	19,200	Total Eat Rate	0.241	Invalid Outcome Rate	0.1%
Total Cells	64	Winners Rate	0.759	Valid Trajectories	19,185
Total Trials	3,200	Median TTE	14.8 steps	Invalid Trajectories	15
Risk Horizon	19	RMST (steps)	14.8
		Total Winners	14,566

Table 1: Dataset overview: counts, survival metrics, and quality.

Panel of Figure 2 — Figure 2: (a) Overall Event Distribution. The plot shows the probability of an agent eating the marshmallow for the first time at each step, aggregated across all conditions; (b) Forest Plot of OR from the discrete-time hazard model. Values less than 1 indicate a reduction in the hazard of eating, while values greater than 1 indicate an increase.

Panel of Figure 3 — Figure 3: KM survival curves by agent characteristics, shown with social context. The y-axis represents the proportion of agents still waiting.

Figure 4: Social Influence on Tool Usage (Broadcast vs. Isolated). The plot shows the mean number of questions asked per step.

Panel of Figure 5 — Figure 5: Ablation Completion Rates: Broadcast vs. Isolated and Survival Curves for all Ablation Conditions

AI Agent Benchmarks

AI agent benchmarks measure how reliably autonomous systems plan, use tools, interact, and complete tasks across controlled and real-world settings.

See all 1,441 papers

Evaluation of Multi-Turn Consistency in LLM Agents: Survival Analysis and Failure-Rationale Taxonomy
Aug 26, 2026Igor Bogdanov, Olga Manakina, Chung-Horng LungBehavioral Consistency of LLM Agents LLM Agent Reliability
Systems and Computer Engineering Carleton University Ottawa, ON, Canada · Department of Cognitive Science Carleton University Ottawa, ON, Canada
Ask Now, Use Later: Benchmarking the Proactivity Gap in Long-Lived LLM Agents
May 27, 2026Bin Wu, Guanyun Zou, Bingbing Wang +2Proactive Assistance LLM Agent Memory
Beijing University of Posts and Telecommunications · Nanjing University of Aeronautics and Astronautics · Harbin Institute of Technology (Shenzhen) +1
Interaction Theater: A case of LLM Agents Interacting at Scale
Feb 23, 2026Sarath Shekkizhar, Adam EarleMulti-Agent Systems Multi-Agent Communication
Salesforce AI Research

Aug 26, 2026J/K move · Enter open · S save

Evaluation of Multi-Turn Consistency in LLM Agents: Survival Analysis and Failure-Rationale Taxonomy

Igor Bogdanov, Olga Manakina, Chung-Horng Lung

Systems and Computer Engineering Carleton University Ottawa, ON, Canada · Department of Cognitive Science Carleton University Ottawa, ON, Canada

Behavioral Consistency of LLM Agents LLM Agent Reliability Reasoning Consistency in Language Models Long-Horizon Agent Evaluation

Open paper AskPDF arXiv

Large language model (LLM) agents may perform well on isolated tasks yet drift into inconsistency over extended interaction. We evaluate temporal consistency in a controlled 20-step multi-agent setting inspired by delayed-gratification studies. At each step, an agent chooses between continuing to delay a reward or claiming it immediately (terminating the episode). Across a full-factorial manipulation of social visibility (private vs public), persona stressors, and deliberation policy, we run 84,540 trajectories spanning 8 model families. Treating the first reward-claim as a time-to-event outcome, we estimate Kaplan-Meier survival curves and fit discrete-time hazard regression to quantify how experimental factors shift failure risk over time. Then, to analyze rationales and language patterns associated with failure, we build a seven-category taxonomy from 13,780 deliberation traces from agents who choose to terminate the episode, using an LLM-assisted labeling paired with human audit (

κ=0.83

). Rationale profiles change systematically with time and context: early failures are more impulse-driven, later failures more fatigue- and cost-benefit-framed, while public settings increase norm-oriented justifications. We also find a deliberation-inconsistency association: among failures, longer deliberation correlates with higher rates of intra-rationale contradiction (simultaneous pro-delay and pro-claim statements), challenging the assumption that more reasoning text implies greater consistency. Together, the survival and rationale analyses reveal distinct temporal reliability regimes and model-specific "failure fingerprints", offering an evaluation lens for diagnosing inconsistency in multi-turn agent behavior.

Abstract

AI Agent Benchmarks

Explore similar work

Evaluation of Multi-Turn Consistency in LLM Agents: Survival Analysis and Failure-Rationale Taxonomy

Ask Now, Use Later: Benchmarking the Proactivity Gap in Long-Lived LLM Agents

Interaction Theater: A case of LLM Agents Interacting at Scale