Tag: #reinforcement-learning9 results
BAO uses agentic RL to train proactive LLM agents balancing performance and user engagement. Combines behavior enhancement with regularization to align with user expectations. Outperforms baselines on UserRL benchmarks.
ARC introduces a reinforcement learning policy to dynamically configure LLM-based agent systems per query, selecting optimal workflows, tools, and prompts. It outperforms fixed templates on reasoning and tool-augmented QA benchmarks. The approach boosts accuracy by up to 25% while cutting token and runtime costs.
DRPO tackles model collapse in off-policy generative recommendation via optimistic distributionally robust optimization. Proves hard filtering recovers high-quality data from noisy logs. Achieves SOTA on mixed-quality benchmarks.
PiT-PO uses reinforcement learning to evolve LLMs for symbolic regression, enforcing physical validity and parsimony. It treats LLMs as adaptive generators updated by search feedback. Achieves SOTA on benchmarks and discovers novel turbulence models.
OmniVL-Guard unifies vision-language forgery detection and grounding with balanced RL. Self-Evolving CoT and ARSPO tackle difficulty bias. Excels in zero-shot generalization.
NSAM learns symbolic models and action masks automatically during DRL to avoid infeasible actions. It integrates symbolic reasoning with deep policy optimization mutually. Evaluations show improved sample efficiency and fewer violations.
Online Causal Kalman Filtering models IS ratios as evolving latent states for stable RL in LLMs. Smooths noise while preserving token structure. Superior on math reasoning datasets.
Introduces Blockwise Advantage Estimation for GRPO in structured generations, assigning per-objective advantages to avoid interference. Uses Outcome-Conditioned Baseline to estimate advantages without nested rollouts. Competitive on math tasks with uncertainty estimation.
Proposes causal reward shaping from offline data for continuous RL under confounders. Derives tight value bounds via causal Bellman equation for PBRS. Outperforms SAC on benchmarks.