AdamO: A Collapse-Suppressed Optimizer for Offline RL

• URL: http://arxiv.org/abs/2605.01968v1
• Date: Sun, 03 May 2026 16:53:29 GMT
• Title: AdamO: A Collapse-Suppressed Optimizer for Offline RL
• Abstract summary: offline reinforcement learning can fail spectacularly when bootstrapped temporal-difference (TD) updates amplify their own errors, driving the critic toward extreme and unusable Q-values.<n>We model offline TD learning as a feedback system and analyze Adam-based critic updates.<n>We prove that this design theoretically guarantees worst-case task safety and preserves Adam's continuous-time dissipative dynamics.
• Score: 12.482937306811026
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Offline reinforcement learning (RL) can fail spectacularly when bootstrapped temporal-difference (TD) updates amplify their own errors, driving the critic toward extreme and unusable Q-values. A key counterintuitive insight of this work is that collapse is not only a property of the backup rule or network architecture: optimizer dynamics themselves can directly trigger or suppress instability. From a control-theoretic viewpoint, we model offline TD learning as a feedback system and analyze Adam-based critic updates. This yields a necessary and sufficient condition for stability of the induced local update dynamics: within the regime we analyze, these dynamics are stable if and only if the spectral radius of the corresponding update operator is strictly below one. Further analysis suggests that standard Adam updates can inadvertently distort the parameter geometry, motivating explicit orthogonality constraints to prevent TD error amplification. To this end, we propose AdamO, an Adam-based optimizer with a decoupled orthogonality correction regulated by a strict task-alignment budget. We prove that this design theoretically guarantees worst-case task safety and preserves Adam's continuous-time dissipative dynamics. Empirically, AdamO is broadly compatible with diverse offline RL baselines, improving stability and returns across a broad suite of benchmarks.

Related papers

• CAST: Critique-Aware Supervision for Training Reliable Long-Horizon Tool-Calling Agents [66.9575676266637]
  Large language model (LLM) agents are increasingly deployed in long-horizon, interactive, and stateful environments.<n>In these settings, a single wrong action, such as refunding the wrong purchase, can cause irreversible task failure and must be intercepted before execution.<n>We develop CAST, a critique-aware training framework that converts sparse task outcomes into action-level supervision for critique learning and policy optimization.
  arXiv  Detail & Related papers  (2026-08-31T01:59:51Z)
• Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models [55.788110316999166]
  We compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against instruction-tuned baselines.<n>We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions.<n>These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence.
  arXiv  Detail & Related papers  (2026-06-09T16:14:27Z)
• StableGrad: Backward Scale Control without Batch Normalization [1.1145952934885128]
  Training very deep neural networks requires controlling the propagation of magnitudes across depth.<n>Modern architectures often mitigate this problem through Batch Normalization, residual connections, or other normalization layers.<n>We propose StableGrad, an alternative mechanism that corrects layer-wise weightgradient imbalances without modifying the forward model.
  arXiv  Detail & Related papers  (2026-05-19T13:49:30Z)
• Decoupled Orthogonal Dynamics: Regularization for Deep Network Optimizers [7.170491058218013]
  We argue that magnitude and direction play distinct roles and should be decoupled in dynamics.<n>We propose Orthogonal Dynamics Decoupling and instantiate it as AdamO: an SGD-style update handles the onedimensional norm control, while Adam's adaptive preconditioning is confined to the tangential subspace.<n> Experiments on vision and language tasks show that AdamO improves generalization and stability over AdamW without introducing additional complex constraints.
  arXiv  Detail & Related papers  (2026-02-04T23:41:51Z)
• ADARL: Adaptive Low-Rank Structures for Robust Policy Learning under Uncertainty [28.291179179647795]
  We propose textbfAdaptive Rank Representation (AdaRL), a bi-level optimization framework that improves robustness.<n>At the lower level, AdaRL performs policy optimization under fixed-rank constraints with dynamics sampled from a Wasserstein ball around a centroid model.<n>At the upper level, it adaptively adjusts the rank to balance the bias--variance trade-off, projecting policy parameters onto a low-rank manifold.
  arXiv  Detail & Related papers  (2025-10-13T20:05:34Z)
• Beyond 'Aha!': Toward Systematic Meta-Abilities Alignment in Large Reasoning Models [86.88657425848547]
  Large reasoning models (LRMs) already possess a latent capacity for long chain-of-thought reasoning.<n>We explicitly align models with three meta-abilities: deduction, induction, and abduction, using automatically generated, self-verifiable tasks.<n>Our three stage-pipeline individual alignment, parameter-space merging, and domain-specific reinforcement learning, boosts performance by over 10% relative to instruction-tuned baselines.
  arXiv  Detail & Related papers  (2025-05-15T17:58:33Z)
• Retrieval is Not Enough: Enhancing RAG Reasoning through Test-Time Critique and Optimization [58.390885294401066]
  Retrieval-augmented generation (RAG) has become a widely adopted paradigm for enabling knowledge-grounded large language models (LLMs)<n>RAG pipelines often fail to ensure that model reasoning remains consistent with the evidence retrieved, leading to factual inconsistencies or unsupported conclusions.<n>We propose AlignRAG, a novel iterative framework grounded in Critique-Driven Alignment (CDA)<n>We introduce AlignRAG-auto, an autonomous variant that dynamically terminates refinement, removing the need to pre-specify the number of critique iterations.
  arXiv  Detail & Related papers  (2025-04-21T04:56:47Z)
• Large Continual Instruction Assistant [59.585544987096974]
  Continual Instruction Tuning (CIT) is adopted to instruct Large Models to follow human intent data by data.<n>Existing update gradient would heavily destroy the performance on previous datasets during CIT process.<n>We propose a general continual instruction tuning framework to address the challenge.
  arXiv  Detail & Related papers  (2024-10-08T11:24:59Z)
• Understanding, Predicting and Better Resolving Q-Value Divergence in Offline-RL [86.0987896274354]
  We first identify a fundamental pattern, self-excitation, as the primary cause of Q-value estimation divergence in offline RL. We then propose a novel Self-Excite Eigenvalue Measure (SEEM) metric to measure the evolving property of Q-network at training. For the first time, our theory can reliably decide whether the training will diverge at an early stage.
  arXiv  Detail & Related papers  (2023-10-06T17:57:44Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.