Reinforcement learning from verifiable rewards (RLVR) is an emerging paradigm for improving the reasoning ability of large language models. However, standard on-policy training discards rollout experiences after a single update, leading to computational inefficiency and instability. While prior work on RL has highlighted the benefits of reusing past experience, the role of experience characteristics in shaping learning dynamics of large reasoning models remains underexplored. In this paper, we are the first to investigate what makes a reasoning experience valuable and identify rollout correctness and entropy as effective indicators of experience value. Based on these insights, we propose ExGRPO (Experiential Group Relative Policy Optimization), a framework that organizes and prioritizes valuable experiences, and employs a mixed-policy objective to balance exploration with experience exploitation. Experiments on five backbone models (1.5B-8B parameters) show that ExGRPO consistently improves reasoning performance on mathematical/general benchmarks, with an average gain of +3.5/7.6 points over on-policy RLVR. Moreover, ExGRPO stabilizes training on both stronger and weaker models where on-policy methods fail. These results highlight principled experience management as a key ingredient for efficient and scalable RLVR.

Fine-grained visual reasoning remains a core challenge for multimodal large language models (MLLMs). The recently introduced ReasonMap highlights this gap by showing that even advanced MLLMs struggle with spatial reasoning in structured and information-rich settings such as transit maps, a task of clear practical and scientific importance. However, standard reinforcement learning (RL) on such tasks is impeded by sparse rewards and unstable optimization. To address this, we first construct ReasonMap-Plus, an extended dataset that introduces dense reward signals through Visual Question Answering (VQA) tasks, enabling effective cold-start training of fine-grained visual understanding skills. Next, we propose RewardMap, a multi-stage RL framework designed to improve both visual understanding and reasoning capabilities of MLLMs. RewardMap incorporates two key designs. First, we introduce a difficulty-aware reward design that incorporates detail rewards, directly tackling the sparse rewards while providing richer supervision. Second, we propose a multi-stage RL scheme that bootstraps training from simple perception to complex reasoning tasks, offering a more effective cold-start strategy than conventional Supervised Fine-Tuning (SFT). Experiments on ReasonMap and ReasonMap-Plus demonstrate that each component of RewardMap contributes to consistent performance gains, while their combination yields the best results. Moreover, models trained with RewardMap achieve an average improvement of 3.47% across 6 benchmarks spanning spatial reasoning, fine-grained visual reasoning, and general tasks beyond transit maps, underscoring enhanced visual understanding and reasoning capabilities.

Reinforcement learning with human-annotated data has boosted chain-of-thought reasoning in large reasoning models, but these gains come at high costs in labeled data while faltering on harder tasks. A natural next step is experience-driven learning, where models improve without curated labels by adapting to unlabeled data. We introduce RESTRAIN (REinforcement learning with Self-restraint), a self-penalizing RL framework that converts the absence of gold labels into a useful learning signal. Instead of overcommitting to spurious majority votes, RESTRAIN exploits signals from the model's entire answer distribution: penalizing overconfident rollouts and low-consistency examples while preserving promising reasoning chains. The self-penalization mechanism integrates seamlessly into policy optimization methods such as GRPO, enabling continual self-improvement without supervision. On challenging reasoning benchmarks, RESTRAIN delivers large gains using only unlabeled data. With Qwen3-4B-Base and OctoThinker Hybrid-8B-Base, it improves Pass@1 by up to +140.7 percent on AIME25, +36.2 percent on MMLU_STEM, and +19.6 percent on GPQA-Diamond, nearly matching gold-label training while using no gold labels. These results demonstrate that RESTRAIN establishes a scalable path toward stronger reasoning without gold labels.

We introduce Policy Gradient Guidance (PGG), a simple extension of classifier-free guidance from diffusion models to classical policy gradient methods. PGG augments the policy gradient with an unconditional branch and interpolates conditional and unconditional branches, yielding a test-time control knob that modulates behavior without retraining. We provide a theoretical derivation showing that the additional normalization term vanishes under advantage estimation, leading to a clean guided policy gradient update. Empirically, we evaluate PGG on discrete and continuous control benchmarks. We find that conditioning dropout-central to diffusion guidance-offers gains in simple discrete tasks and low sample regimes, but dropout destabilizes continuous control. Training with modestly larger guidance ($γ>1$) consistently improves stability, sample efficiency, and controllability. Our results show that guidance, previously confined to diffusion policies, can be adapted to standard on-policy methods, opening new directions for controllable online reinforcement learning.

Understanding how subjective experience arises from information processing remains a central challenge in neuroscience, cognitive science, and AI research. The Modular Consciousness Theory (MCT) proposes a biologically grounded and computationally explicit framework in which consciousness is a discrete sequence of Integrated Informational States (IISs). Each IIS is a packet of integrated information tagged with a multidimensional density vector that quantifies informational richness. Its magnitude correlates with subjective intensity, shaping memory, behavior, and continuity of experience. Inputs from body and environment are adaptively filtered, processed by modules (abstraction, narration, evaluation, self-evaluation), and integrated into an IIS. The resulting packet, tagged with its density vector, is transmitted to behavioral readiness, memory, and decision-making modules, closing the loop. This explains why strongly tagged states exert greater influence on long-term memory and action. Unlike Global Workspace Theory, Integrated Information Theory, or Higher-Order Thought, MCT specifies a full computational pipeline producing discrete informational units with quantifiable internal structure. Subjectivity is reframed as a correlate of the density-tagging signal with functional consequences. MCT generates testable predictions, such as stress enhancing memory encoding, and provides a naturalistic blueprint for both biological and artificial architectures. Consciousness, in this view, is not an irreducible essence but an evolvable, quantifiable, and constructible feature of complex information processing.

We reframe and operationalise adversarial inverse reinforcement learning (IRL) to large language model reasoning, learning a dense, token-level reward model for process supervision directly from expert demonstrations rather than imitating style via supervised fine-tuning. The learned reasoning reward serves two complementary roles: (i) it provides step-level feedback to optimise a reasoning policy during training; and (ii) it functions at inference as a critic to rerank sampled traces under fixed compute budgets. We demonstrate that our approach prioritises correctness over surface form, yielding scores that correlate with eventual answer validity and enabling interpretable localisation of errors within a trace. Empirically, on GSM8K with Llama3 and Qwen2.5 backbones, we demonstrate: (i) dense reasoning rewards can be used as a learning signal to elicit reasoning, and (ii) predictive performance is improved from reward-guided reranking (notably for Llama-based policies). By unifying training signals, inference-time selection, and token-level diagnostics into a single reasoning reward, this work suggests reusable process-level rewards with broad potential to enhance multi-step reasoning in language models.

Object-Goal Navigation (ObjectNav) is a critical component toward deploying mobile robots in everyday, uncontrolled environments such as homes, schools, and workplaces. In this context, a robot must locate target objects in previously unseen environments using only its onboard perception. Success requires the integration of semantic understanding, spatial reasoning, and long-horizon planning, which is a combination that remains extremely challenging. While reinforcement learning (RL) has become the dominant paradigm, progress has spanned a wide range of design choices, yet the field still lacks a unifying analysis to determine which components truly drive performance. In this work, we conduct a large-scale empirical study of modular RL-based ObjectNav systems, decomposing them into three key components: perception, policy, and test-time enhancement. Through extensive controlled experiments, we isolate the contribution of each and uncover clear trends: perception quality and test-time strategies are decisive drivers of performance, whereas policy improvements with current methods yield only marginal gains. Building on these insights, we propose practical design guidelines and demonstrate an enhanced modular system that surpasses State-of-the-Art (SotA) methods by 6.6% on SPL and by a 2.7% success rate. We also introduce a human baseline under identical conditions, where experts achieve an average 98% success, underscoring the gap between RL agents and human-level navigation. Our study not only sets the SotA performance but also provides principled guidance for future ObjectNav development and evaluation. Recent advances in computer vision and deep learning have inspired growing interest in interdisciplinary applications that bridge perception, reasoning, and control, especially in robotics. Among these, vision-based navigation has emerged as a foundational capability for autonomous mobile agents. A key benchmark in this domain is Object-Goal Navigation (ObjectNav), where a robot must navigate to an instance of a specified object category in an unseen environment, relying solely on its onboard sensors. This task is both practically important and technically challenging: it requires semantic understanding, spatial reasoning, and long-horizon planning. Among many approaches, Reinforcement Learning (RL) has become a dominant paradigm for ObjectNav, offering a structured framework to learn directly through trial-and-error and showing steady progress across various benchmarks. While end-to-end RL policies are common, modular RL approaches have shown greater robustness and improved generalization.

Distributionally robust reinforcement learning (DRRL) focuses on designing policies that achieve good performance under model uncertainties. In particular, we are interested in maximizing the worst-case long-term discounted reward, where the data for RL comes from a nominal model while the deployed environment can deviate from the nominal model within a prescribed uncertainty set. Existing convergence guarantees for robust temporal-difference (TD) learning for policy evaluation are limited to tabular MDPs or are dependent on restrictive discount-factor assumptions when function approximation is used. We present the first robust TD learning with linear function approximation, where robustness is measured with respect to the total-variation distance and Wasserstein-l distance uncertainty set. Additionally, our algorithm is both model-free and does not require generative access to the MDP. Our algorithm combines a two-time-scale stochastic-approximation update with an outer-loop target-network update. We establish an $\tilde{O}(1/ε^2)$ sample complexity to obtain an $ε$-accurate value estimate. Our results close a key gap between the empirical success of robust RL algorithms and the non-asymptotic guarantees enjoyed by their non-robust counterparts. The key ideas in the paper also extend in a relatively straightforward fashion to robust Q-learning with function approximation.

Vision-Language-Action (VLA) models aim to unify perception, language understanding, and action generation, offering strong cross-task and cross-scene generalization with broad impact on embodied AI. However, current VLA models often lack explicit step-by-step reasoning, instead emitting final actions without considering affordance constraints or geometric relations. Their post-training pipelines also rarely reinforce reasoning quality, relying primarily on supervised fine-tuning with weak reward design. To address these challenges, we present VLA-R1, a reasoning-enhanced VLA that integrates Reinforcement Learning from Verifiable Rewards (RLVR) with Group Relative Policy Optimization (GRPO) to systematically optimize both reasoning and execution. Specifically, we design an RLVR-based post-training strategy with verifiable rewards for region alignment, trajectory consistency, and output formatting, thereby strengthening reasoning robustness and execution accuracy. Moreover, we develop VLA-CoT-13K, a high-quality dataset that provides chain-of-thought supervision explicitly aligned with affordance and trajectory annotations. Furthermore, extensive evaluations on in-domain, out-of-domain, simulation, and real-robot platforms demonstrate that VLA-R1 achieves superior generalization and real-world performance compared to prior VLA methods. We plan to release the model, code, and dataset following the publication of this work. Code: https://github.com/GigaAI-research/VLA-R1. Website: https://gigaai-research.github.io/VLA-R1.

LLM-based multi-agent systems excel at planning, tool use, and role coordination, but their openness and interaction complexity also expose them to jailbreak, prompt-injection, and adversarial collaboration. Existing defenses fall into two lines: (i) self-verification that asks each agent to pre-filter unsafe instructions before execution, and (ii) external guard modules that police behaviors. The former often underperforms because a standalone agent lacks sufficient capacity to detect cross-agent unsafe chains and delegation-induced risks; the latter increases system overhead and creates a single-point-of-failure-once compromised, system-wide safety collapses, and adding more guards worsens cost and complexity. To solve these challenges, we propose AdvEvo-MARL, a co-evolutionary multi-agent reinforcement learning framework that internalizes safety into task agents. Rather than relying on external guards, AdvEvo-MARL jointly optimizes attackers (which synthesize evolving jailbreak prompts) and defenders (task agents trained to both accomplish their duties and resist attacks) in adversarial learning environments. To stabilize learning and foster cooperation, we introduce a public baseline for advantage estimation: agents within the same functional group share a group-level mean-return baseline, enabling lower-variance updates and stronger intra-group coordination. Across representative attack scenarios, AdvEvo-MARL consistently keeps attack-success rate (ASR) below 20%, whereas baselines reach up to 38.33%, while preserving-and sometimes improving-task accuracy (up to +3.67% on reasoning tasks). These results show that safety and utility can be jointly improved without relying on extra guard agents or added system overhead.