Liquid AI
Liquid AI's LFM2.5-350M: Big Performance, Tiny Model
Liquid AI releases a 350M parameter model trained on 28 trillion tokens with scaled reinforcement learning, challenging assumptions about what compact models can achieve.
Reinforcement Learning
LLM evaluation
REAL: New RL Method Improves LLM Judge Accuracy for AI Evaluation
Researchers introduce REAL, a regression-aware reinforcement learning framework that trains LLM judges to produce more accurate evaluations by optimizing for numerical precision rather than classification.
AI Safety
Research: LLM Safety Training Survives RL Optimization
New research examines whether safety guardrails in large language models remain intact when agents are optimized for helpfulness through reinforcement learning.
LLM
Reinforcement-Aware Knowledge Distillation Advances LLM Reasoning
New research combines reinforcement learning with knowledge distillation to improve how smaller language models learn complex reasoning from larger teacher models.
LLM Agents
Tool-R0: LLM Agents That Learn Tool Use Without Training Data
New research introduces Tool-R0, a framework enabling LLM agents to autonomously learn tool usage through self-evolution, eliminating the need for curated training datasets while achieving state-of-the-art performance.
AI Agents
MIRA: Combining Memory Systems with RL for Smarter AI Agents
New research introduces MIRA, a framework that integrates memory architectures with reinforcement learning while minimizing expensive LLM calls, advancing efficient autonomous agent design.
Agentic AI
Proxy State Evaluation: Scaling Verifiable Rewards for AI Agents
New research proposes proxy state-based evaluation for multi-turn tool-calling LLM agents, addressing the challenge of scalable reward verification in complex agentic workflows.
LLM
Error-Localized Policy Optimization: A New Approach to LLM Tool R
New research introduces ELPO, a training method that teaches LLMs to learn from irrecoverable errors in tool-integrated reasoning chains, improving agent capabilities.
LLM Agents
Agent-Omit: Teaching LLMs to Think More Efficiently
New research introduces Agent-Omit, a reinforcement learning framework that trains LLM agents to selectively omit unnecessary reasoning steps and observations, dramatically improving computational efficiency.
LLM Infrastructure
PROTEUS: Lagrangian RL Optimizes Multi-LLM Routing for SLAs
New research introduces PROTEUS, a reinforcement learning framework using Lagrangian optimization to intelligently route requests across multiple LLMs while meeting strict service level agreements.
LLM Research
Process-Supervised RL: Precise Error Penalization Boosts LLM Reas
New research introduces a method to preserve correct reasoning steps while penalizing errors, improving LLM performance through more nuanced reinforcement learning credit assignment.
Voice AI
RLAIF Breakthrough: Optimizing Spoken AI Without Human Feedback
New research demonstrates how reinforcement learning from AI feedback can optimize spoken dialogue systems using multiple LLM evaluators, reducing dependency on costly human annotations.