A new arXiv preprint introduces 474 executable games to test LLM interactive reasoning. The benchmark reveals that counterfactual revision and necessity judgment cause much larger performance drops than contextual perturbations.
Key facts
- 474 executable games in the benchmark
- Five difficulty levels per game configuration
- Counterfactual revision causes larger drops than perturbations
- Submitted to arXiv on May 26, 2026
A team of researchers led by Mingyuan Fan, Weiguang Han, and Daixin Wang has released a multi-turn interactive framework for evaluating LLM reasoning, instantiated as a benchmark of 474 executable games. Each game requires the model to receive only task rules, then issue targeted queries to a hidden environment, integrate partial observations over time, and decide when to submit a final answer [According to Evaluating Interactive Reasoning in Large Language Models].
The benchmark evaluates models across five difficulty levels, each with fixed configuration search spaces. Beyond standard success rate and interaction efficiency, the framework measures contextual robustness under controlled perturbations and metacognitive adaptation through counterfactual revision and necessity judgment.
Results show the benchmark is highly discriminative, exposing large differences not only in success rate but also in interaction efficiency across frontier LLMs. Critically, the authors empirically show that contextual perturbations cause moderate but consistent declines, whereas counterfactual revision and necessity judgment lead to much larger drops. This suggests current models lack robust metacognitive capabilities — the ability to revise beliefs when counterfactual evidence contradicts prior observations.
The unique take here is that static benchmarks like SWE-Bench or GSM8K miss a fundamental failure mode: LLMs can't effectively update beliefs through active interaction. The 474-game setup mirrors real-world agent scenarios where models must query databases, APIs, or environments rather than solve isolated problems. The large gap between standard interaction and counterfactual revision suggests agentic AI systems may fail catastrophically when assumptions are violated.
Key Takeaways
- New 474-game benchmark reveals LLMs fail on counterfactual reasoning, with larger drops than contextual perturbations.
- Highlights metacognitive gaps in agentic AI.
How the Benchmark Works
The framework implements Algorithm 1 (Interactive Protocol) where each game has a hidden state. The model issues queries, receives partial observations, and must balance information gathering against answering. Table 1 in the paper breaks down games by data structure and reasoning type, while Table 2 reports overall performance on the clean interactive reasoning backbone — measuring success rate, average turns over successful episodes, and efficiency defined as Success Rate / Avg. Turns.
The authors evaluated a broad set of frontier LLMs but did not disclose specific model names or scores in the abstract. The paper is available on arXiv under cs.AI, submitted May 26, 2026.
Implications for Agentic AI
This benchmark arrives as Meta, OpenAI, and Anthropic race to deploy agentic AI systems. Meta recently mandated 65-80% of developer code be AI-generated by mid-2026, and internal AI agents have already triggered security incidents [As previously reported]. The finding that counterfactual reasoning causes large drops suggests these systems may struggle when production environments deviate from training conditions — a common real-world scenario.
What to watch
Watch for follow-up papers that disclose specific model scores and per-model breakdowns on counterfactual revision tasks. Also track whether Meta, OpenAI, or Anthropic adopt this benchmark for internal agent evaluation — a strong signal of its industry relevance.
