verifiers

This paper from Salesforce audits public RL environments for terminal agents, finding significant defects, and introduces RIVER, a training recipe that filters defective environments and penalizes repetitive behavior to enhance model performance.

The author shares a novel use case for Jev in building custom verifiers for AI agent harnesses, enabling scalable test-time compute by combining System One and System Two models.

The tweet expresses skepticism about powerful entities acting in concert and emphasizes the need to verify the verifiers, predicting their growing influence in the future.

The article discusses the problem of inconsistent AI agent workflows in engineering teams and introduces OmniNode, a tool that standardizes agent work through explicit criteria and verification to enhance system reliability.

VerMem is a framework for unified memory management in LLM agents, using local and global verifiers with reinforcement learning to jointly control long-term and short-term memory. It outperforms strong baselines across five benchmarks with improved efficiency-performance trade-offs.

Prime Intellect released verifiers v1 and prime-rl 0.7.0, an RL training tool with full support for verifiers, multiple algorithms like GRPO and OPD, and performance improvements.

Prime Intellect Lab is out of beta, offering a platform to train models with support for various architectures and modalities, enabling self-improving agents.

Andrej Karpathy frames LLM training as text, conversations, and environments; Prime Intellect's Verifiers is an open-source framework for building and sharing RL environments for LLMs, released under MIT license, with a hub of 2500+ environments.

A technical tutorial on building a reinforcement learning environment for LLMs using the open-source Verifiers library, with Othello as a working example.

Omar highlights the growing value of building LLM verifiers and judges for agentic coding, while Mira Murati shares that Bridgewater partnered with TinkerAPI to fine-tune a model for financial analysis.

A multi-tweet analysis of ~15 agentic-loop papers concludes that the verifier, not the model, is the key predictor of success, with examples showing that robust, non-gamable checks (e.g., compilers, tests, verifiable rewards) dramatically improve performance, while failures stem from lack of such verifiers or gaming vulnerabilities.

Emphasizes the importance of verifiers for LLM-based agents, noting that out-of-distribution tasks cause failures, and suggests tuning custom verifiers.

The user is working on implementing reasoning training with verifiers using Unsloth and TRL, reporting progress on locally generating GRPO-like rollouts with a small SLM and a tiny RM, and promises a video soon.

Researchers propose an adversarial hacker-fixer loop using LLM agents to automatically patch brittle verifiers in agent benchmarks, reducing attack success rates from 62% to 0% on KernelBench and demonstrating that weaker defenders can neutralize much stronger attackers.

Tweet highlighting work on making verifiers cheaper for scaling evaluations and reinforcement learning, by researchers from Harvey.

A study by LangChain and Harvey explores methods to reduce the cost of verifying legal agent outputs by batching criteria evaluations and using open models, achieving order-of-magnitude cost savings while maintaining near-frontier performance.

This paper investigates reward hacking in rubric-based reinforcement learning, analyzing the divergence between training verifiers and evaluation metrics. It introduces a diagnostic for the 'self-internalization gap' and demonstrates that stronger verification reduces but does not eliminate reward hacking.

AgentV-RL introduces an Agentic Verifier framework that enhances reward modeling through bidirectional verification with forward and backward agents augmented with tools, achieving 25.2% improvement over state-of-the-art ORMs. The approach addresses error propagation and grounding issues in verifiers for complex reasoning tasks through multi-turn deliberative processes combined with reinforcement learning.

OpenAI trained a system using verifiers to solve grade school math word problems with 90% of child-level accuracy, nearly doubling fine-tuned GPT-3 performance. The approach addresses language models' weakness in multistep reasoning by training verifiers to evaluate candidate solutions and select the best one.