STS2-Bench: A Benchmark for Long-Term Decisions
MrBoxChen · reddit · 2026-07-12
The author proposed a new benchmark, STS2-Bench, using Slay the Spire 2 to test models' long-term decision-making capabilities.
Core Concept
Instead of answering isolated questions, the model must continuously make decisions throughout an entire game:
- Read constantly changing states
- Balance short-term gains with long-term survival
- Choose cards, routes, and resources under uncertainty
- Adjust strategies based on outcomes after each round
Experiments and Findings
- The author evaluated 7 model configurations within the same framework
- 5.6Sol performed surprisingly well on this sequential decision-making task
- The author explicitly stated that this shouldn't be viewed as a general intelligence ranking, but it might reveal signals hidden from traditional single-turn benchmarks
Feedback Requested
- Whether this game task is suitable as a proxy metric for planning and long-range reasoning
- What control variables or baselines are needed to make the evaluation more credible
- What other games or environments would be suitable for this type of benchmark
Related event: STS2-Bench Uses Slay the Spire 2 to Test Long-Horizon Decisions(2 posts)→
More from Research
- LightMTP: lightweight latent multi-token prediction boosts LLM planning in pretraining — gdm3000 · 2026-10-27
- EMNLP study: AI literary translation reads fine but varies more, even within one book — tuvllms · 2026-10-27
- Lab monkeys at $30-80K each: a 4-year breeding cycle is the real preclinical drug bottleneck — zebird0 · 2026-10-27
- OpenAI Foundation funds UNC tumor antigen project, hiring senior experimental scientist — iskander · 2026-10-27
- Steering attention heads toward honesty: up moral patienthood, down psychopathy — edelwax · 2026-10-27
- Are transcoders overkill? New paper argues plain neurons suffice for circuit analysis — burny_tech · 2026-10-27