My RL Digest

• Sample-Routed Policy Optimization (SRPO) routes correct samples to GRPO's reward-aligned reinforcement and failed ones to SDPO's logit-level...

Quantum RL made accessible – no hardware needed. Install via pip install qrl-qai==1.0.0 in Colab for Gymnasium+PennyLane+PyTorch environments.

-...

New RL Papers

• 🔥 GrandCode: GrandCode achieves grandmaster level in competitive programming via agentic reinforcement learning and placed first...

Humanoid robotics trend: Sim-to-real RL breakthroughs enable production-ready feats.

• KAIST's lower-body platform sprints at 12 km/h, moonwalks, and...

Self-Distilled RLVR paper released by @_akhaliq: https://t.co/5oucSjKaJs https://t.co/CwH09W9j5F. Fresh RL breakthrough for vision tasks.

Key advances pushing LLM agents toward real-world coding prowess:

• GrandCode multi-agent RL system tops Codeforces live contests (Rounds 1087-1089),...

Algorithmic Advances

• 🔥 Policy Gradient RL with Separated Knowledge: Paper presents policy gradient reinforcement learning with separated...

AgentHazard is a new benchmark for evaluating harmful behavior in computer-use agents, spotlighting risks in real-world AI deployments. Join the discussion on the paper page.

Emerging trend in LLM+RL agent efficiency:

• RLCF breakthrough: Uses community feedback on high/low-citation papers to teach AI scientific taste,...

Heracles bridges precise motion tracking and generative synthesis for humanoid robots:

• Acts as intermediary using flow matching for real-time...

Policy gradient RL breakthrough: Replaces state-values—previously used as reusable parameters for behavior knowledge—with separated knowledge.

Key evolution in LLM robot control:

• Modular past: LLMs planned tasks (e.g., SayCan), but separate networks handled motor signals, limiting...

RL's shortcut problem undermines reliable model reasoning. Court's reasoning hinges on understanding how these models think, exposing critical risks for real-world RL deployment.

New RARLDA framework applies RL to microgrid dispatch, integrating CVaR for weather risks like max wind speed, rainfall, and temperature in coastal...

Algorithmic RL Breakthroughs

• 🔥 DeepMind AlphaEvolve: Google DeepMind proposes AlphaEvolve, an LLM-powered evolutionary coding agent that...

Q-CQL fuses conservative Q-learning with quantum computing for robust decisions in noisy data environments.

• Conservative edge: Avoids risky...

Emerging RL-LLM synergy drives breakthroughs:

• Azure RFT elevates LLM reasoning for enterprise agents like retail tools and DocuSign contracts, using...

MACE launches three Gymnasium-compatible environments for stock trading, margin trading, and portfolio optimization.

• Realistic modeling: Nonlinear...

Humanoid RL surges with sim-to-real breakthroughs:

• Full pipeline sims enable RL training without hardware, from sensors to control.
• Deep RL boosts...

Hardware Breakthroughs

• 🔥 Stable Memristor for RL: New memristor design with built-in oxygen gradient produces slow, stable conductance changes,...