Evaluation
Evaluation of Multi-Turn Consistency in LLM Agents
Terminal success rates conceal distinct temporal reliability regimes. Survival analysis and failure-rationale taxonomy reveal when models fail and how they justify abandoning a goal.
- Venue
- Workshop paperICLR 2026 Workshop on LLM Logical Reasoning
- Role
- First author · methodology · evaluation infrastructure · analysis
- Themes
- Evaluation
Schematic survival curves for three regimes (near-flat hazard, early spike, and bimodal) drawn over early, middle, and late failure bands. A row showing the dominant pre-failure rationale moving from impulse to cost–benefit to fatigue. And a qualitative heatmap of rationale mix per model labelled failure fingerprint. Curves are illustrative, not measured values.
- 84,540
- trajectories
- eight model families, 64 conditions per family
- 14,025
- pre-failure traces
- reasoning captured immediately before goal abandonment
- 13,780
- labeled rationales
- seven-category taxonomy
- κ = 0.83
- human-audit agreement
- on rationale labels
01
The problem
Most agent benchmarks report a terminal success rate. Two models with the same success rate can be very different systems: one fails early and at random, the other holds steady and then collapses late. A single number hides this, and it also hides what the model was thinking when it gave up.
This work treats goal abandonment as an event in time and borrows the tools of survival analysis (hazard curves, time-to-failure distributions) to compare models. It pairs that with a taxonomy of the rationales models produce immediately before failing, so the temporal pattern and the narrative pattern can be read together.
The environment is a long-horizon survival micro-benchmark in which an agent must repeatedly resist an immediate reward to preserve a longer-term goal, with configurable social exposure and deliberation requirements. Horizon, reward structure, and prompt templates were held constant within each condition so that differences are attributable to model and condition.
02
System and method
Schematic of the pipeline and the three hazard regimes observed. Trajectories are run to failure or horizon. Survival curves and hazard functions are estimated per model and condition. The reasoning trace preceding each failure is classified into a seven-category taxonomy and audited by humans. The joint distribution forms a per-model failure fingerprint. Curves are illustrative, not measured values.
03
Experimental design
- Scale
- 84,540 trajectories across eight model families
- Conditions
- 64 conditions per family, varying social exposure, mandatory deliberation, and related factors
- Event definition
- Goal abandonment (taking the immediate reward) at a given turn
- Temporal analysis
- Survival curves and hazard estimates per model and condition
- Rationale corpus
- 14,025 pre-failure traces · 13,780 labeled with a seven-category taxonomy
- Validation
- Human audit of labels with κ = 0.83 agreement
- Additional measures
- Argument density and intra-rationale contradiction within failure traces
04
Main findings
- 01
Models exhibit near-flat, early-spike, and bimodal hazard regimes
Hazard curves separate the tested models into distinct regimes: some fail at a roughly constant rate, some concentrate failures in the first turns, and some show both an early spike and a late rise. Terminal success rates do not reveal which regime a model is in.
Why it mattersChoosing a model for a long-horizon task requires knowing when it fails, not just how often.
- 02
Early failures are impulse-driven, later ones are cost–benefit or fatigue-framed
Rationale categories shift with time. Failures in the first turns are dominated by impulse-like justifications. Failures after sustained persistence are more often framed as a cost–benefit calculation or as fatigue.
Why it mattersMitigations should be stage-specific: guarding against impulse early is a different problem from preventing late-stage rationalized abandonment.
- 03
Public settings produce social-contagion rationales
When agents could observe others' choices, a distinct rationale category appeared in which the model justified abandonment by reference to what others had done.
Why it mattersMulti-agent deployments introduce failure narratives that do not exist in isolation and should be evaluated explicitly.
- 04
Mandatory deliberation changes narratives without improving reliability
Forcing the model to reason before acting altered the distribution of rationales but did not materially improve survival. Greater argument density within a rationale was associated with more intra-rationale contradiction among failures.
Why it mattersLonger reasoning is not a reliability intervention by itself. It can produce more elaborate justifications for the same failure.
05
What I built
The engineering behind the result. Every number above was produced by systems I designed and implemented.
Benchmark environment
Text-based multi-agent gym modelled as an MDP/POMDP with configurable social exposure, personas, and deliberation.
Experiment runner
Parallel orchestration of 84,540 trajectories across eight model families with resumable, seeded runs.
Model adapters
Multi-provider connectors with retries, rate-limit handling, and token accounting.
Trace capture
Structured logging of the reasoning trace preceding every failure event.
Labeling pipeline
Taxonomy-driven classification of 13,780 rationales with a human-audit protocol.
Survival analysis
Time-to-event estimation and hazard visualization per model and condition.
Contradiction analysis
Argument-density and intra-rationale contradiction measures over failure traces.
06
Limitations
07
Artifacts
- paperPaper (OpenReview)Evaluation of Multi-Turn Consistency in LLM Agents: Survival Analysis and Failure-Rationale Taxonomy. ICLR 2026 Workshop on LLM Logical Reasoning. (opens in a new tab)
- posterCompanion benchmark (NeurIPS 2025 MTI Workshop)Delay-of-Gratification as a Multi-Agent Survival Micro-Benchmark: the gym underlying this evaluation. (opens in a new tab)
08
Related research
- ArchitectureContext, Reasoning, and Hierarchy in Compound LLM AgentsWhat an agent sees can matter more than how long it deliberates.Open project
- AdaptationFORGE: Self-Evolving Agent Memory With No Weight UpdatesPrompt-only reflection becomes more reliable when useful discoveries are selected and broadcast across a population.Open project
Contact
Building reliable AI systems requires both research and engineering.
I’m interested in research engineering, applied research, agent infrastructure, evaluation, reliability, interpretability, and research-to-production work.