Gradient Equilibrium shown equivalent to Blackwell Approachability in online learning
A new arXiv preprint proves that gradient equilibrium (GEQ), a recently introduced online optimization framework generalizing first-order stationarity, is algorithmically equivalent to Blackwell approachability. The equivalence implies GEQ is also equivalent to regret minimization and calibration, resolving an open question about GEQ's place in the online learning landscape. The reductions are efficient and allow transfer of refined guarantees like optimism and strong adaptivity from regret minimization to GEQ, with applications including online conformal prediction.
Related guides (1)
Related events (8)
Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning
This paper introduces Equilibrium Reasoners (EqR), a framework that formalizes test-time compute scaling through learned task-conditioned attractors in latent space, where stable fixed points correspond to valid solutions. EqR scales along two axes—depth (more iterations) and breadth (aggregating stochastic trajectories)—without requiring external verifiers or task-specific priors. On Sudoku-Extreme, unrolling up to 40,000 equivalent layers boosts accuracy from 2.6% (feedforward baseline) to over 99%. The work provides a mechanistic lens for understanding why iterative latent models generalize beyond memorized patterns.
Global Convergence Theory for Wasserstein Policy Gradient in Entropy-Regularized RL
This paper establishes the first global convergence theory for Wasserstein Policy Gradient (WPG), a continuous-control RL optimization method that uses optimal-transport geometry over action distributions. The authors show that the Bellman recursion structure of entropy-regularized RL induces a Polyak–Łojasiewicz (PL) geometry that substitutes for classical convexity, enabling global convergence analysis. Key technical contributions include a statewise KL representation of the soft Bellman residual, a Bellman resolvent identity linking value improvement to relative Fisher information, and a uniform log-Sobolev inequality for the evolving Gibbs policy family. The result yields geometric contraction up to discretization bias, providing theoretical grounding for WPG in continuous-action settings.
Equivalence between Policy Gradients and Soft Q-Learning
OpenAI published a research result establishing a formal equivalence between policy gradient methods and soft Q-learning, two major families of reinforcement learning algorithms. The work shows that under entropy regularization, these approaches are mathematically equivalent, unifying previously separate lines of RL research. This has implications for algorithm design, theoretical understanding, and the development of hybrid RL methods.
General Preference Reinforcement Learning (GPRL): Bridging Online RL and Preference Optimization for Open-Ended Tasks
GPRL proposes a new alignment framework that replaces scalar reward models with a General Preference Model (GPM) embedding responses into k skew-symmetric subspaces to capture multi-dimensional, intransitivity-aware preferences. The method computes per-dimension group-relative advantages, normalizes across axes, and uses a closed-loop drift monitor to detect and correct single-axis reward hacking during training. Starting from Llama-3-8B-Instruct, GPRL achieves a 56.51% length-controlled win rate on AlpacaEval 2.0 and outperforms SimPO and SPPO on Arena-Hard, MT-Bench, and WildBench. The work directly addresses the gap between verifiable-reward online RL (strong on math/code) and preference optimization (strong on open-ended tasks).
Kolmogorov Regression lifts diffusion policies to Cameron-Martin space for robust long-horizon control
Researchers introduce a backward Kolmogorov equation framework that reformulates diffusion policy training as a deterministic boundary-value PDE problem in Cameron-Martin space, replacing stochastic score matching. The approach uses a precision-weighted Cameron-Martin loss and a Kolmogorov residual as an inference-time failure detector, yielding convergence guarantees tied to kernel effective rank rather than action dimension. Validation on the PushT manipulation benchmark shows 17% improvement in episode reward and 67.6% reduction in inter-step drift; a 6-station manufacturing scheduling task shows 28.4% lower RMSE than LSTM baselines and 96% reduction in deadlock events via Hamilton-Jacobi reachability certification.
GGRO: Gradient-Guided Reward Optimization for inference-time LLM alignment
Researchers introduce Gradient-Guided Reward Optimization (GGRO), an inference-time alignment method that uses gradient signals from a reward model to inject 'nudging tokens' at high-uncertainty decoding steps, rather than relying on sampling-intensive re-ranking approaches like Best-of-N. The method monitors token-level entropy to detect distribution drift and steers generation trajectories directly, claiming improved robustness to reward hacking with minimal computational overhead. Experiments show gains across safety, helpfulness, and reasoning benchmarks compared to standard inference-time alignment baselines.
Tight Convergence Theory for Error Feedback Algorithms in Distributed Optimization
This paper provides tight convergence analyses for two major error-feedback algorithms—classic Error Feedback (EF) and Error Feedback 21 (EF21)—used to mitigate communication bottlenecks in distributed learning. The authors identify optimal step-size choices and construct tailored Lyapunov functions for each method, yielding guarantees that hold independently of the number of agents and recover the best known single-agent bounds. The work clarifies the relative performance of these gradient compression variants, which has remained poorly understood despite widespread use.
Open problem paper questions whether AdamW converges under heavy-tailed gradient noise
A preprint from arXiv frames as an open problem whether AdamW, the dominant optimizer for LLM pretraining, can achieve rigorous convergence guarantees under heavy-tailed stochastic gradient noise. The authors note that sign-based optimizers like Lion and Muon already have sharp heavy-tailed convergence rates, while AdamW's second-moment accumulator may create a fundamental obstruction by hiding large gradients. The paper proves a positive weighted-metric benchmark and introduces a corridor lower-bound mechanism to characterize the potential failure mode.