Multi-Turn Reflective Masking Elicits Reasoning in Mask Diffusion Models

While reasoning on autoregressive (AR) models is often performed by chain-of-thought reasoning and reflection, their refinement of previous outputs still relies on fully sequential generation, even when only local edits are needed. In contrast, the masking mechanism in Mask Diffusion Models (MDMs) naturally supports explicit local edits on previous outputs, allowing selective refinement without discarding previous answers and generating another from scratch. While this property more closely aligns with how humans correct mistakes by iterative local refinement, existing MDMs do not support multi-turn masking and denoising. We propose Reflective Masking (RM), which elicits such an intrinsic reasoning capability in MDMs via lightweight post-training. RM provides a native test-time scaling, where an MDM iteratively revisits and revises its prior outputs based on evolving context. To exploit insights from previous turns like AR reasoning, we further introduce History Reference, a parameter-free mechanism that leverages intermediate denoising states during revision. Our approach requires no architectural changes and is easily applicable to existing MDMs. Across diverse tasks and modalities, including text generation, Sudoku, and image editing, Reflective Masking consistently outperforms standard masking-based baselines and demonstrates strong generality, positioning RM as a fundamental primitive for reasoning on MDMs.

Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization

Reward hacking is usually studied after it becomes visible, once a model earns high proxy reward while failing the intended task. We instead study what proxy RL teaches before that failure appears. We introduce Proxy Reward Internalization and Mechanistic Exploitation (PRIME), a learned capability to assess task correctness, predict proxy acceptance, and reason about exploitable proxy--gold gaps. In coding RL environments with exploitable pytest rewards, we measure PRIME through chain-of-thought monitoring, direct probes, and activation-level concept vectors. We find that PRIME emerges in a staged sequence before sustained reward hacking, and that its current direct-probe score forecasts later hack onset and severity even when the visible hack rate is still low. PRIME also adapts when the evaluator changes, retargeting to whichever proxy--gold gap remains rewarded and persisting when gold reward suppresses overt hacking, and ablating its activation directions reduces hacking. Across checkpoints, in-domain PRIME tracks out-of-domain misalignment. Together these results suggest that exploitable proxy RL amplifies a proxy-internalization capability upstream of visible hacking, making PRIME a candidate early-warning signal for broader alignment risk.

Learning Geometric Representations from Videos for Spatial Intelligent Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) excel at 2D semantic understanding but lack intrinsic 3D awareness, resulting in representations that fail to maintain geometric and spatial consistency across video frames. Given the scarcity of large-scale 3D data, we present GeoVR, a novel framework that learns geometric representations using purely 2D video sequences. This approach effectively restructures the semantic latent space within MLLMs to unlock spatial intelligence. Rather than employing superficial feature mixing, GeoVR reshapes the internal representations of the MLLM by distilling geometry knowledge from pre-trained 3D foundation models. This is accomplished through a multi-objective learning strategy driven by four complementary geometric targets: (1) estimating inter-frame camera poses to embed varying viewpoint dynamics, (2) regressing dense depth maps to anchor physical distances, (3) predicting a metric scale factor for real-world calibration, and (4) distilling multi-scale 3D features to align the intermediate feature space. Guided by these explicit physical and geometric constraints, the model's internal representations naturally develop strong 3D awareness. Extensive experiments on spatial reasoning benchmarks demonstrate that GeoVR achieves state-of-the-art performance, establishing a new paradigm for endowing foundation models with spatial intelligence.

Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimization

Process reward models (PRMs) provide fine-grained reward signals along the reasoning process, but training reliable PRMs often requires step annotations or heavy verification pipelines, making them expensive to scale and refresh during online RL. Implicit PRMs mitigate this cost by learning decomposable token- or step-level rewards from trajectory-level outcome labels. However, they suffer from a train-inference mismatch: training only constrains a sequence-level aggregate, whereas inference requires token-level scores to reflect local step quality. As a result, token-level credits are weakly identified and may fail to faithfully reflect which reasoning steps are actually correct. This unreliability undermines a key promise of implicit PRMs: scoring many candidate tokens. In practice, noisy per-token advantages may systematically reinforce incorrect continuations. We address this problem with a novel Implicit Prefix-Value Reward Model (IPVRM), which directly learns a prefix-conditioned value function estimating the probability of eventual correctness, and derives step signals via temporal-difference (TD) differences. IPVRM substantially improves step-verification F1 on ProcessBench. Building on these calibrated prefix values, we further propose Distribution-Level RL (DistRL), which computes TD advantages for both sampled tokens and high-probability candidate tokens, enabling dense counterfactual updates without additional rollouts. While DistRL offers limited gains when powered by miscalibrated implicit rewards, it consistently improves downstream reasoning once paired with IPVRM.

Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding

3D Visual Grounding (3D-VG) aims to localize objects in 3D scenes via natural language descriptions. While recent advancements leveraging Vision-Language Models (VLMs) have explored zero-shot possibilities, they typically suffer from a static workflow relying on preprocessed 3D point clouds, essentially degrading grounding into proposal matching. To bypass this reliance, our core motivation is to decouple the task: leveraging 2D VLMs to resolve complex spatial semantics, while relying on deterministic multi-view geometry to instantiate the 3D structure. Driven by this insight, we propose "Think, Act, Build (TAB)", a dynamic agentic framework that reformulates 3D-VG tasks as a generative 2D-to-3D reconstruction paradigm operating directly on raw RGB-D streams. Specifically, guided by a specialized 3D-VG skill, our VLM agent dynamically invokes visual tools to track and reconstruct the target across 2D frames. Crucially, to overcome the multi-view coverage deficit caused by strict VLM semantic tracking, we introduce the Semantic-Anchored Geometric Expansion, a mechanism that first anchors the target in a reference video clip and then leverages multi-view geometry to propagate its spatial location across unobserved frames. This enables the agent to "Build" the target's 3D representation by aggregating these multi-view features via camera parameters, directly mapping 2D visual cues to 3D coordinates. Furthermore, to ensure rigorous assessment, we identify flaws such as reference ambiguity and category errors in existing benchmarks and manually refine the incorrect queries. Extensive experiments on ScanRefer and Nr3D demonstrate that our framework, relying entirely on open-source models, significantly outperforms previous zero-shot methods and even surpasses fully supervised baselines.

Incentivizing Temporal-Awareness in Egocentric Video Understanding Models

Multimodal large language models (MLLMs) have recently shown strong performance in visual understanding, yet they often lack temporal awareness, particularly in egocentric settings where reasoning depends on the correct ordering and evolution of events. This deficiency stems in part from training objectives that fail to explicitly reward temporal reasoning and instead rely on frame-level spatial shortcuts. To address this limitation, we propose Temporal Global Policy Optimization (TGPO), a reinforcement learning with verifiable rewards (RLVR) algorithm designed to incentivize temporal awareness in MLLMs. TGPO contrasts model outputs generated from temporally ordered versus shuffled video frames to derive calibrated, globally normalized reward signals that explicitly favor temporally coherent reasoning. Integrated with GRPO and GSPO, TGPO supports cold-start RL training and effectively suppresses spatial shortcut behaviors learned by existing MLLMs. Experiments across five egocentric video benchmarks demonstrate that TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches. Our results suggest that TGPO offers a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding.

IR$^3$: Contrastive Inverse Reinforcement Learning for Interpretable Detection and Mitigation of Reward Hacking

Reinforcement Learning from Human Feedback (RLHF) enables powerful LLM alignment but can introduce reward hacking - models exploit spurious correlations in proxy rewards without genuine alignment. Compounding this, the objectives internalized during RLHF remain opaque, making hacking behaviors difficult to detect or correct. We introduce IR3 (Interpretable Reward Reconstruction and Rectification), a framework that reverse-engineers, interprets, and surgically repairs the implicit objectives driving RLHF-tuned models. We propose Contrastive Inverse Reinforcement Learning (C-IRL), which reconstructs the implicit reward function by contrasting paired responses from post-alignment and baseline policies to explain behavioral shifts during RLHF. We then decompose the reconstructed reward via sparse autoencoders into interpretable features, enabling identification of hacking signatures through contribution analysis. Finally, we propose mitigation strategies - clean reward optimization, adversarial shaping, constrained optimization, and feature-guided distillation - that target problematic features while preserving beneficial alignment. Experiments across multiple reward model configurations show that IR3 achieves 0.89 correlation with ground-truth rewards, identifies hacking features with over 90% precision, and significantly reduces hacking behaviors while maintaining capabilities within 3% of the original model.

StagePilot: A Deep Reinforcement Learning Agent for Stage-Controlled Cybergrooming Simulation

Cybergrooming is an evolving threat to youth, necessitating proactive educational interventions. We propose StagePilot, an offline RL-based dialogue agent that simulates the stage-wise progression of grooming behaviors for prevention training. StagePilot selects conversational stages using a composite reward that balances user sentiment and goal proximity, with transitions constrained to adjacent stages for realism and interpretability. We evaluate StagePilot through LLM-based simulations, measuring stage completion, dialogue efficiency, and emotional engagement. Results show that StagePilot generates realistic and coherent conversations aligned with grooming dynamics. Among tested methods, the IQL+AWAC agent achieves the best balance between strategic planning and emotional coherence, reaching the final stage up to 43% more frequently than baselines while maintaining over 70% sentiment alignment.

Adversarial Reward Auditing for Active Detection and Mitigation of Reward Hacking

Reinforcement Learning from Human Feedback (RLHF) remains vulnerable to reward hacking, where models exploit spurious correlations in learned reward models to achieve high scores while violating human intent. Existing mitigations rely on static defenses that cannot adapt to novel exploitation strategies. We propose Adversarial Reward Auditing (ARA), a framework that reconceptualizes reward hacking as a dynamic, competitive game. ARA operates in two stages: first, a Hacker policy discovers reward model vulnerabilities while an Auditor learns to detect exploitation from latent representations; second, Auditor-Guided RLHF (AG-RLHF) gates reward signals to penalize detected hacking, transforming reward hacking from an unobservable failure into a measurable, controllable signal. Experiments across three hacking scenarios demonstrate that ARA achieves the best alignment-utility tradeoff among all baselines: reducing sycophancy to near-SFT levels while improving helpfulness, decreasing verbosity while achieving the highest ROUGE-L, and suppressing code gaming while improving Pass@1. Beyond single-domain evaluation, we show that reward hacking, detection, and mitigation all generalize across domains -- a Hacker trained on code gaming exhibits increased sycophancy despite no reward for this behavior, and an Auditor trained on one domain effectively suppresses exploitation in others, enabling efficient multi-domain defense with a single model.

TokenSeek: Memory Efficient Fine Tuning via Instance-Aware Token Ditching

Fine tuning has been regarded as a de facto approach for adapting large language models (LLMs) to downstream tasks, but the high training memory consumption inherited from LLMs makes this process inefficient. Among existing memory efficient approaches, activation-related optimization has proven particularly effective, as activations consistently dominate overall memory consumption. Although prior arts offer various activation optimization strategies, their data-agnostic nature ultimately results in ineffective and unstable fine tuning. In this paper, we propose TokenSeek, a universal plugin solution for various transformer-based models through instance-aware token seeking and ditching, achieving significant fine-tuning memory savings (e.g., requiring only 14.8% of the memory on Llama3.2 1B) with on-par or even better performance. Furthermore, our interpretable token seeking process reveals the underlying reasons for its effectiveness, offering valuable insights for future research on token efficiency. Homepage: https://runjia.tech/iclr_tokenseek/