README.md

metadata

title: README
emoji: 👁
colorFrom: green
colorTo: purple
sdk: static
pinned: false

Fine-grain evaluation & Large Reasoning Models that fails in reasoning due to reasoning rigidity.
ConditionedMath (AIME & MATH500) · PuzzleTrivial · Zero-shot pipelines

---

📜 Why ReasoningTrap?

Current RL-tuned Reasoning LLMs excel at producing answers but often ignore explicit user constraints.
ReasoningTrap surfaces these failure modes with carefully crafted, conditioned problems.

• Modified from Famous MATH Reasoning Benchmark – AIME & MATH500 problems altered with minimal constraints to divert reasoning paths.
• Puzzles Trivialized by Subtle Modifications - Well-known puzzles where a small change transforms a challenging problem into a trivial one.
• Plug-and-play – evaluate any 🤗 Transformers model with vLLM in simple instructions.