CoSineVerifier: Tool Augmented Answer Verification for Computation Oriented Scientific Questions. 📖 Introduction We present CoSineVerifier Tool , a compact, tool augmented verifier for C omputation O riented S c I e N tific s E nario answer verification. It equips LLM reasoning with external tools—e.g., Python interpreter,—to accurately verify answers in computation oriented scenarios such as algebraic equivalence and physical constant alignment. We also release the CoSineVerifier series: efficient single token labeling verifiers with performance comparable to CoSineVerifier Tool. Our methods achieve state of the art results on VerifyBench and SCI VerifyBench. They also show clear improvements on RLVR tasks over other verification methods. Picture 1: RLVR performance with different verification methods. 🌟 Key Features: Tool augmented verification for computation oriented scientific scenarios : CoSineVerifier Tool evaluates multi step reasoning in math and science where correctness depends on precise intermediate calculations. It also normalizes answers to a common dimension/unit for consistent numeric checks. It integrates: 🧮Python Interpreter: executes snippets to validate alge…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy