每天知道
你的领域
出了什么

Distribution Matching Distillation (DMD) trains a few-step student from the difference between separately estimated target and student scores, so it must keep an auxiliary diffusion model fitted to the student's evolving distribution at extra memory and computation cost. We introduce DMAD, Distribution Matching as Adversarial Distillation, which recasts distribution matching as classification and learns the required log-density ratios directly. Two discriminator heads on a shared backbone distinguish real data and teacher samples from the student's, and linear losses on their logits train the student without auxiliary score fitting. We prove that at the discriminator optimum these losses recover the distribution-matching gradient underlying DMD, through the classical identity linking discriminator logits to log-density ratios. We further introduce gap-based reweighting, which adapts teacher supervision across noise levels from the real-data head's empirical logit gap between real and teacher samples. DMAD reaches a Fréchet Inception Distance (FID) of 1.04 with one-step generation on ImageNet-64x64, 14.47 with four-step SDXL on COCO-10K, and a VBench total score of 85.15 with four-step Wan2.1-T2V-14B, the best values among the compared few-step methods and the multi-step teachers. On MiniMax-H3-33B, our four-step student achieves overall human preference rates of 79.1% over DMD2 and 84.6% over rCM for joint audio-video generation, excluding ties. Our code, models and demos are available at https://yzmblog.github.io/projects/DMAD.

Verifier-free reinforcement learning with probability-based rewards offers a promising way to train LLMs on general reasoning tasks where external verifiers are unavailable. Yet the reliability of these rewards, especially in long-horizon reasoning, remains underexplored. This work identifies a length-dependent failure mode of probability rewards, which we call the Posterior Concentration Phenomenon (PCP). We show that the probability of a reference answer conditioned on a reasoning trace often collapses to a low-variance interval as the trace becomes lengthy. This phenomenon results in nearly indistinguishable rewards, which, under GRPO-based settings, makes probability-based policy optimization unstable and inefficient. Motivated by this, we propose Reinforcement Learning with Concentration-aware Posterior Rewards (RLCPR), a verifier-free RL framework to explicitly account for PCP for better optimization stability and token efficiency. It has two components: uncertainty-aware data sampling, which reduces concentration-prone rollouts before generation, and concentration-aware regularization, which penalizes unnecessarily long traces when posterior rewards collapse. Extensive experiments show that, alongside higher token efficiency, RLCPR outperforms the state-of-the-art verifier-free RL baseline by up to 4.0% on six of seven benchmarks, including general-domain and mathematical reasoning challenges.

Agentic reinforcement learning (RL) has emerged as a powerful approach for training large language model agents on multi-step tasks, yet reliance on terminal outcome rewards creates two credit-assignment problems, particularly in long-horizon tasks. First, same-outcome rollout groups provide no learning signal from terminal rewards. Second, terminal rewards provide only trajectory-wide feedback, making it difficult to identify which decisions caused a failure. Recent work supplements terminal rewards with finer-grained information from trajectory analysis, such as natural-language reflections on intermediate decisions and errors. However, natural-language diagnoses are difficult to use directly for credit assignment: their error claims may be unreliable, and they do not quantify how much each error should affect learning. We propose Self-Diagnosis-guided Terminal Credit Redistribution (FAULT), which turns diagnosed errors into explicit step-level credit anchored by terminal outcomes. FAULT checks diagnostic evidence and learns relative error costs from task outcomes. During training, the policy and self-diagnoser co-evolve, while error costs are updated online from recent outcomes. On ALFWorld, FAULT recovers learning signals from same-outcome groups, reaching 95% signal coverage versus 41% for GRPO and 72% for GiGPO, while better localizing credit to specific error steps. Across two model scales, FAULT delivers strong. improvements on the long-horizon ALFWorld and WebShop tasks while remaining competitive on short-horizon Search-based QA.

Quadrupedal locomotion requires balancing conflicting objectives such as command tracking, stability, and energy efficiency, yet conventional reinforcement learning (RL) hardcodes these priorities into a fixed scalar reward at training time. We present PROMO (Preference-Conditioned Multi-Objective Reinforcement Learning), a semantic multi-objective approach that makes this trade-off an explicit runtime input to a single locomotion policy. PROMO conditions the policy on deployment facing preferences while keeping embodiment-specific locomotion priors fixed, thereby separating operator intent from reward shaping terms required for viable gait generation. Compared with fixed-objective controllers, multi-objective baselines, and independently trained specialists, PROMO achieves objective specialization and robustness from a single deployable policy. Across 100 sampled preferences in simulation, 67 behaviors are non-dominated under exact Pareto dominance, with a mean preference-objective correlation of 0.843, demonstrating broad Pareto coverage and predictable preference response. The same policy transfers zero-shot to a Unitree Go2, where preference changes alone reduce specific energy by up to 30.4%, position error by 38.7%, and peak body-attitude deviation by 59.0% relative to the balanced preference. These results establish preference-conditioned multi-objective RL as a practical runtime interface for adaptive legged locomotion, extending its role beyond offline Pareto-set construction. Open-source code and videos are available at https://amrmousa.com/promo/.

Asynchronous reinforcement learning (RL) improves the efficiency of large language model post-training but introduces stale rollouts generated by earlier policies. Theoretical understanding of how this staleness affects convergence and how to mitigate its impact remains limited. We derive a convergence bound for GRPO-style algorithms that explicitly characterizes the tradeoff between the gradient estimator's second moment and bias. For trajectory-level importance-weighted estimators, our analysis shows that once the second moment is uniformly controlled, delay enters the bound through the bias introduced by clipping or rescaling. Guided by this insight, we propose a novel group mass capping GRPO (GMC-GRPO) method, which minimizes a ratio-based bias bound within a class of weighted estimators sharing a common second-moment guarantee. We establish convergence guarantees for asynchronous GMC-GRPO and show that, compared with TIC-GRPO, it improves the threshold dependence of the fourth-order delay term from $O(ε^{-4})$ to $O(ε^{-2})$ as $ε\to0$, where $1+ε$ is the ratio threshold. Under local policy overlap, the delay-dependent term decreases as $G^{-2/5}$ after tuning the step size, where $G$ is the group size. For fixed behavior and current policies, the bias introduced by group rescaling also vanishes as $G\to\infty$, whereas the bias from trajectory-wise clipping can persist. Experiments across Qwen3 models and reasoning benchmarks demonstrate improved robustness to stale rollouts, with GMC-GRPO achieving the best performance among stable baselines under large rollout delays.

Outcome-based reinforcement learning can train language models to forecast real-world events, but prior forecasting work either freezes research context before training or deploys agentic research only at test time, so the skill of gathering evidence is never shaped by the reward. We introduce an agentic forecasting environment, dataset, and harness built from 2,100+ resolved Polymarket questions; the agent acquires its own context at rollout time (web search, page reading, and financial time series, all restricted by layered leak filtering to information published before each question's cutoff), and we train Qwen3.5-35B-A3B (3B active parameters) on it with single-epoch GRPO under a Brier-score reward. Training changes how the agent interacts with information: calibration improves 30-40%, and search attempts fall from 3.8 to 2.25 per rollout as evidence discipline is learned. Evaluated in an identical harness against four frontier models, the trained policy also finishes ahead of every frontier model tested at evidence-based forecasting, including Claude Opus 4.5 (soft-Brier 0.254 vs. 0.256, n=265), at about 5% of the inference cost, and its margin is widest on the hardest questions, the ones the crowd itself had not decided. We release the environment, dataset, and per-rollout records as a reusable harness for temporal forecasting agents.

Science teachers frequently search for documentary excerpts not by describing what appears on screen, but by querying the abstract concepts they intend to teach. This use case exposes a limitation of existing language-based video moment retrieval methods, which typically assume that queries describe observable events, whereas instructional search requires retrieving concrete visual phenomena that instantiate an underlying scientific principle. We study this setting as concept-to-example video retrieval, an abstract-needle-in-a-haystack problem where compact curriculum concepts must be grounded in temporally sparse documentary evidence. To bridge this abstraction gap, we propose Concept-Driven Domain Adaptation (CDDA), a three-stage framework for adapting two-tower vision-language models to concept-level retrieval. CDDA treats concepts as intermediate semantic anchors: it first structures the textual embedding space with textbook and teacher-handbook example-concept pairs, then transfers this concept-aware geometry to documentary visuals under a frozen visual encoder, and finally jointly adapts both encoders with sparse visual concept supervision. From a geometric perspective, this staged alignment reduces text-concept and vision-concept angular gaps, thereby encouraging concept-level adaptation while preserving the pretrained model's concrete image description alignment. On a curated middle-school physics retrieval benchmark, CDDA achieves stronger pedagogically oriented concept retrieval than several competitive multimodal baselines, including Qwen3-VL-Embedding-2B, while maintaining concrete image-text matching after adaptation.

Language model (LM) alignment broadly aims to perturb a given LM $Q$ into an aligned LM $q$ such that i) the outputs produced by $q$ and $Q$ are 'close' in probability, ii) $q$ has a higher expected reward than $Q$. Two common techniques for LM alignment are: KL-constrained RL, which requires knowledge of the LM distribution and is computationally expensive, and the best-of-$n$ algorithm, which requires only sampling from the LM. The work of Yang et al. established asymptotic closeness between the distributions produced by the two alignment methods for an $m$--length i.i.d. token sequence output by the LM, in the limit as $m$ increases to infinity. However, the i.i.d. assumption is not representative of practical LMs, whose output sequences often have memory. In this paper, we extend the asymptotic closeness result to the case when the $m$--length token sequence outputted by the LM is Markovian. Further, for finite-length output sequences — particularly, when $m=1$ — we provide a complete characterization of LM distributions and reward functions for which the KL-divergence between the distributions produced by the two alignment methods is zero — a question first posed in Yang et al.

The rapid growth of video-based social media has increased users' exposure to harmful content, creating a need for reliable automated video safety detection. Although recent Vision-Language Models (VLMs) show strong video understanding capabilities, existing harmful video detection systems face two key limitations: they typically reduce safety detection to binary classification, overlooking the inherently multi-label nature of unsafe videos, and they rely on static training objectives that do not support controllable precision-recall trade-offs, though the desired operating point may vary across moderation pipelines and unsafe categories. To address these gaps, we propose Adaptive Tversky Policy Optimization (ATPO), a reinforcement learning framework for Multi-label Video Safety Detection (Multi-VSD). ATPO introduces the Adaptive Tversky Reward (ATR), which dynamically adjusts false-positive and false-negative penalties during training to enable controllable precision-recall trade-offs. Experiments on SafeWatch-Bench and XD-Violence show that ATPO substantially improves multi-label performance, increasing the Jaccard Index from 40.66 to 75.44 on SafeWatch-Bench-Real. Moreover, ATR enables reliable steering of the precision-recall operating point, supporting deployment scenarios with heterogeneous policy requirements. Code and checkpoints are provided at https://bruceyg.github.io/ATPO-project-page/ .

We introduce an approach for collection-grounded B-roll sequence generation. Given a user's video collection, a directive given in natural language, and a target duration, the goal is to produce a multi-shot sequence that complements the user's primary footage (A-roll) while preserving the collection's characters, settings, objects, and style. This task is challenging as one must choose the visual evidence from hours of captured footage that should guide the generation of each shot in the sequence. We address this challenge with MemComposer, a three-stage system that turns raw footage into a structured memory with visual references (characters, settings, objects, and style) and uses it to plan, retrieve, and generate grounded B-roll sequences. First, in a one-time offline stage, MemComposer constructs an entity-centric memory from raw video. Second, it uses the memory and user directive to plan a grounded sequence and retrieve conditioning frames for each shot. Third, it iteratively generates and critiques the sequence to enforce identity, setting, and sequence-level consistency. We evaluate MemComposer in a user preference study along two dimensions: prompt adherence and visual alignment to the user's collection. Against an ungrounded text-to-video planner, MemComposer wins 60.0% of prompt-adherence and 92.8% of visual-alignment comparisons, showing the grounding benefit of collection memory and reference retrieval. Against retrieval-only sequences assembled from captured footage, MemComposer wins 94.5% of prompt-adherence comparisons, showing the value of generating missing shots, while retrieval-only sequences are preferred for visual alignment in 58.2% of comparisons.

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the mathematical reasoning capabilities of large language models. Recent work introduces search into RLVR rollouts to increase trajectory diversity, but diversity alone does not ensure that the search-induced rollout policy improves upon the current policy. To address this gap, we propose APIVIS, a training-time framework that adapts finite-budget Gumbel search to chunk-level mathematical reasoning. APIVIS combines direct and searched responses within each rollout group, allowing improvements found by search to produce informative relative rewards. It further applies selective supervision to search-improved tokens, preserving a learning signal when uniform group rewards render GRPO ineffective. We show that exact value-guided selection improves the expected verifier reward at each searched state and that this guarantee extends to the complete rollout policy, with a corresponding approximate guarantee under bounded value-estimation error. Experiments on widely recognized mathematical reasoning benchmarks and different model scales demonstrate substantial improvements over competitive search-based methods, validating the effectiveness of APIVIS.

Large language models can produce economic policy responses that sound reasonable, but this does not show that their actions are consistent with economic dynamics. We test this by placing an instruction-tuned language model inside six Snowdrop-backed dynamic stochastic general equilibrium (DSGE) simulators. At each turn, the model observes the economy and a change in economic discourse, selects a bounded policy action, and receives the next simulated state and an economic reward. We implement a common Python interface for repeated rollouts, persistent shocks, state cloning, and rolling-horizon simulation. This setting creates a long-horizon credit-assignment problem. Policy effects may appear several quarters after an action is taken. PPO has a learned value function that can propagate delayed reward to earlier tokens through generalized advantage estimation. GRPO has no learned value function and instead assigns a group-relative advantage from complete rollout returns. It therefore cannot distinguish which earlier turn caused the outcome; if every rollout receives the same return, the normalized advantage is zero. We use PPO as the primary method and GRPO as a matched critic-free baseline. The experiments also test directional semantic signals, reward horizon, trajectory warm starts, cross-simulator transfer, and historically anchored pandemic and monetary-policy shocks. The objective is to judge policy actions by their simulated economic consequences rather than by plausible language alone.

Time series captioning is a fundamental step in time series understanding and can also serve as the bridge between signal and natural language. Supervised fine-tuning (SFT) relies on a larger model's captions and cannot exceed their quality. Reinforcement learning (RL) can, but its rewards were designed for other modalities and other tasks, and they transfer poorly to open-ended generation in the time series domain. We address this by proposing LineupRL, a reinforcement learning with verifiable rewards (RLVR) pipeline whose reward is caption-to-series identification. The reward model is a frozen large language model (LLM) verifier that reads the generated caption and the candidate time series as raw values, never the chart, and must pick the described time series from multiple distractors. Matching is a far lighter demand on the verifier than writing questions or judging a caption, so an off-the-shelf LLM can supply the reward. Across two captioning benchmarks, and on forecasting and reconstruction where the predictor sees only the caption, LineupRL outperforms SFT and RL baselines on every metric. The 3B vision language model (VLM) trained by LineupRL also outperforms, at 1/24 of the parameters, the 72B VLM whose captions the SFT baseline is distilled from. Our case study shows that LineupRL resists reward hacking, and that the captioner it trains both traces the trend and names the values at key points.

Urban environments are shaped by design choices with long-term implications for health, safety, and quality of life, yet evaluating proposed interventions remains costly, time-consuming, and often impractical. Existing geospatial vision methods largely focus on monitoring urban indicators from aerial and street-view imagery, rather than proposing interventions and estimating their effects on such indicators. Moving beyond recognition, we introduce the problem of discovering interventions that improve target indicators for a given aerial or street-view image. We argue that a black-box indicator model, combined with a generative editing model, can serve as an implicit digital twin for testing intervention hypotheses. We present VIDA-Geo , a multi-agent system that explores this intervention space by coordinating segmentation, diffusion-based inpainting, and indicator scoring models to produce interventions that are both perceptually realistic and aligned with real-world policies. We evaluate our system on 8 indicators across aerial and street-view imagery, measuring changes in factors such as perceived safety and greenery. Our approach outperforms existing baselines in many cases, achieving up to 2X higher perceptual quality and policy alignment scores. Finally, our model provides users with multiple candidate interventions, supporting an expert city-planner-in-the-loop workflow.

Flow matching is central to 3D generation, yet in practice its reinforcement learning (RL) methods are largely adapted from 2D visual generation. Representative DPO-, GRPO-, and NFT-style objectives, when applied to negative trajectories, mainly steer predicted velocities away from the corresponding directions without explicitly specifying a target velocity field toward preferred samples. In 3D generation, constrained by pretrained model capabilities, rollout diversity, and reward-distribution complexity, directly applying these RL methods yields limited gains in geometric quality. We introduce a forward-process RL method Dynamic Homing Optimization (DHO), which reformulates negative-trajectory optimization as positive-sample attraction-guided dynamic homing. Specifically, Minimum-Cost Attractive Matching (MAM) assigns each negative sample a distinct positive target, and Time-Aware Dynamic Correction (TDC) then redirects its trajectory toward the target using a remaining-time-aware corrective velocity. Building on asynchronous online DHO, we develop Flow3D-Pro, an image-to-3D geometry generation framework. Experiments show that DHO outperforms representative DPO-, GRPO-, and NFT-style objectives in 3D generation, while Flow3D-Pro produces higher-quality 3D geometry than existing mesh generation methods.

As the use of large language models (LLMs) expands, post-training has become increasingly important for adapting them to downstream tasks. However, obtaining reliable supervision remains costly, especially in domains without reference answers or executable verifiers. LLM-as-a-Judge provides scalable pseudo-rewards for unlabeled responses, but a single point score does not explicitly represent reward uncertainty. This motivates representing pseudo-rewards as conformally calibrated reward ranges. We propose Range-GRPO, a semi-supervised post-training framework that combines limited labeled data with unlabeled prompts. In Group Relative Policy Optimization (GRPO), learning signals depend on relative reward comparisons within each rollout group. The proposed objective compares reward ranges pairwise rather than reducing them to point rewards, allowing interval uncertainty to affect both the magnitude and direction of these signals. Our theoretical analysis characterizes this distinction and shows that the proposed objective recovers the Dr$.$GRPO advantage when all reward ranges collapse to points. Empirically, Range-GRPO achieves the highest in-distribution and out-of-distribution average performance among the evaluated semi-supervised methods while requiring fewer training resources.

Reinforcement learning based post training of diffusion models, such as Denoising Diffusion Policy Optimization (DDPO), optimizes a reverse diffusion process under a reward function. However, current approaches to reward optimizations do so at the cost of diversity and quality. In this paper, we provide better tradeoffs through careful theoretical considerations and method design. We analyze the theoretical framework and mathematically demonstrate that only-latter timestep updates of diffusion model may be harmful for diversity contrary to the conclusions presented in a previous work. Additionally, we propose an incremental Feynman-Kac training based on strong theoretical foundations in order to achieve the best-yet alignment-diversity tradeoffs. We perform extensive experiments and compare our method against related diffusion policy optimization approaches in three different tasks and also provide strong ablations for each component, thus validating strong performance gains in both alignment and diversity.

Training a single LLM agent jointly across diverse interactive environments has attracted increasing attention as a route to generalist agents. Existing curriculum and data-selection strategies often allocate training at the environment level or prioritize local reward-based signals, without explicitly considering relationships between current rollouts across environments for prompt-group selection. Meanwhile, as environments are learned at different rates, all-failure and all-success rollout groups can coexist within a batch, leaving those data without group-relative reward signals. Both challenges highlight limitations of relying solely on scalar rewards in multi-environment RL: they provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. This motivates richer textual feedback, such as rubrics describing rollout behaviours, to guide learning. Beyond rubrics' usage as reward, we repurpose rubrics to guide both online data selection and policy supervision. An LLM judge tags each rollout using a predefined rubric vocabulary shared across environments. The resulting profiles guide the selection of data that aligns with the overall behavioural composition of the mixed-environment batch while limiting overlap with already-selected data. Available positive rubrics (describing desired behaviours) provide privileged context for an on-policy self-distillation teacher, supplying additional token-level supervision, while negative rubrics (describing undesired behaviours) guide subsequent rollout generation away from recurring failure modes. Together, these components form RISED. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment. Rubric-based analysis of RISED can further characterize the behavioural changes accompanying these gains.

Generative modeling has shown strong promise for multivariate time mseries (MTS) forecasting, especially scale to high-dimensional settings. Diffusion-based methods achieve competitive performance but typically require many sampling steps at inference. VAE-based non-iterative forecasting frameworks have therefore emerged as an efficient alternative. Within this line of work, vector quantization (VQ) enables controllable latent space modeling by mapping multivariate series into compact discrete representations. Existing VQ-based forecasting methods, however, typically rely on autoregressive (AR) token generation, which suffers from exposure bias and training-inference mismatch. Flow matching provides an efficient non-autoregressive alternative for latent forecasting, but existing formulations usually initialize transport from a generic Gaussian prior. We instead observe that the trained VQ codebook already captures representative latent prototypes and can thus serve as a more informative prior for flow matching. Based on this insight, we propose ProtoFlow, a forecasting framework that combines vector-quantized autoencoding with Prototype-prior Flow matching. Our method first maps multivariate sequences into a discrete latent space, then constructs a structured prior from the learned codebook, and finally learns a DiT-based rectified flow to transport samples from this prior to future latent representations conditioned on historical observations. By replacing generic noise initialization with a learned prototype prior, ProtoFlow avoids the rollout mismatch of AR token prediction and promotes faster training convergence. Extensive experiments on benchmark datasets show that it consistently achieves superior forecasting performance with efficient inference.

Adverse weather image restoration aims to recover images degraded by rain, haze, snow, and other weather-induced artifacts, thereby improving the robustness of outdoor vision systems. Existing unified restoration models exhibit limited generalization to real-world scenes due to their reliance on synthetic supervision and insufficient semantic constraints. In this paper, we propose a novel student--teacher semi-supervised framework that addresses both challenges. Specifically, we introduce an unreliable database that preserves failed teacher predictions as informative negative samples for contrastive learning, while a reliable database stores high-quality teacher predictions as positive samples. By jointly exploiting reliable pseudo-ground truths and unreliable teacher outputs, the proposed framework learns to enhance desirable restoration characteristics while avoiding common failures. We further propose a phase spectrum-based semantic constraint that replaces computationally expensive text-based supervision with an efficient and naturally aligned semantic prior. An adaptive phase consistency loss is also designed to dynamically balance supervision between the degraded input and teacher pseudo-ground truths according to degradation severity. Extensive experiments on real-world benchmarks demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches in restoration quality and perceptual fidelity while exhibiting stronger generalization to real-world adverse weather conditions.

Foundation models remain vulnerable to spurious correlations and ``Clever Hans'' strategies. Explainable machine learning can find and remove such strategies for classifiers without metadata. For foundation models, no such option exists yet. We propose Disentangled Diffusion Autoencoders (DiDAE). DiDAE wraps a frozen foundation model in a conditional diffusion decoder. A counterfactual is one closed-form edit along a direction of a disentangled dictionary, followed by decoding. The dictionary can be supervised (Procrustes) or unsupervised (Singular Value Decomposition, Sparse Autoencoders). No gradients are needed, so DiDAE is up to 2000 times faster than the state of the art. We evaluate on six datasets, two synthetic and four real-world. In a desiderata-driven benchmark on three of them, its counterfactuals are on par with or better than the state of the art, and they repair downstream classifiers through Counterfactual Knowledge Distillation (CFKD), where they beat metadata-based correction. The same machinery can rank a pretrained dictionary against a trained classifier. It returns the few directions the classifier actually reads, each causally verified by a counterfactual that flips the decision, and repairs the classifier along those a teacher marks spurious. The workflow is plug-and-play in our open-source Peal library we publish alongside the paper. With a public dictionary and a pretrained decoder, all that remains is a cheap linear distillation of the classifier and its own fine-tuning.

Model-based reinforcement learning (MBRL) can exploit simulated experience, but a simulator-to-target shift creates a model-selection problem: correcting the simulator and fitting the target directly can each fail under limited target data. We introduce the Model-Corrected World Model (MC-WM), which separates initial target data into disjoint fit, selection, and calibration partitions and deploys the family with lower standardized calibration risk. A learned confidence signal and deterministic validity predicates weight one-step imagined policy updates without rewriting physical rewards. We evaluate 540 unique reported run cells across three controlled Multi-Joint dynamics with Contact (MuJoCo) shifts; one exact-routing cell was repeated after a pre-deployment artifact gate, giving 541 completed executions.

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher models. We introduce a different form of on-policy self-distillation for MLLMs that provides the teacher with textual, spatially grounded guidance identifying the visual elements relevant to a query. We use procedurally generated scenes with automatically available object identities and spatial coordinates, enabling scalable and annotation-free post-training. The teacher uses this spatial guidance to locate and integrate evidence from multiple relevant image regions, while the student learns to reproduce the resulting behavior from the image and question alone. Our approach consistently improves performance on counting, document and chart understanding benchmarks across multiple models. Importantly, although post-training uses only synthetic scenes, the resulting improvements transfer to real-world perception benchmarks, yielding a 3.23-point gain in average performance across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. These results show that spatially grounded privileged information can induce broader perceptual capabilities through on-policy self-distillation, enabling substantial synthetic-to-real transfer beyond the task and data distribution used for post-training. Project page: https://github.com/sirkosophia/Where-OPD

Extending a text embedding model to new modalities typically degrades text retrieval quality, and existing omni-modal embedders compensate with multi-billion parameters. We present Omni-Embed-Mini, a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into a single shared cosine space without updating any text-side parameter. Our key insight is that the teacher signal requires no separate embedding model: each media sample is paired with a dense cascaded caption, and the teacher target is simply the frozen backbone's own embedding of that caption. Because teacher and student share the same backbone weights, they inhabit byte-identical geometry, and lightweight projectors plus phased LoRA adapters on the modality encoders suffice for alignment. Training combines a Matryoshka SigLIP contrastive loss with an online hybrid hard-negative miner whose negatives sharpen as the encoder improves. The recipe carries over to a 2.3B variant by swapping in a native vision-language backbone. Omni-Embed-Mini-0.9B keeps its text weights bit-identical to the backbone, so training cannot regress text retrieval (49.57 nDCG@10 on MTEB-v2 BEIR-8), while extending it to five additional modalities, and is ~2.7x to 9.5x smaller than every open omni embedder we compare against. The 2.3B variant is competitive with the closed gemini-embedding-2, edging ahead of it on the overall-modality average. Models, code, data and evaluation harness are on our project page: https://omniembed.cvmbzuai.com

Multi-teacher on-policy distillation (MOPD) aims to combine the strengths of RL-trained teachers in a single student, but how teacher signals affect parameter changes remains underexplored. We study Qwen3-1.7B with four domain teachers trained with RL from the same initialization as the student, comparing gradients, optimizer updates, and task learning curves, with additional SmolLM3-3B diagnostics. We find that several factors influence teacher signals. First, loss averaging implicitly weights responses: token averaging favors longer responses, and equalizing domain contributions retains this weighting within domains. Second, Adam's first moment reduces differences in parameter updates: the cosine similarity is 0.83 between teachers and 0.96 between averaging rules, despite differences in raw gradients. Third, BF16 rounding hides small changes: about 97% of FP32 master weights differ from initialization, but only 7--11% of BF16 weights do. Finally, the top-64 intersection KL gradient closely matches Qwen's full-vocabulary gradient, but the effect on task performance depends on averaging: mathematics accuracy is 2.6 points higher than with sampled-token policy-gradient (PG) under response averaging and 2.1 points lower under global token averaging.

Knowledge distillation aims to transfer the factual knowledge of large language models to smaller models for efficient deployment. Yet a teacher may recall a relation in one direction while failing to generate the answer in the reverse direction. Distillation from its generated answers can therefore propagate this directional limitation to the student. The same teacher can nevertheless recognize such an answer by scoring the relation in the direction it knows. We introduce directional label distillation, in which frozen teachers score candidate answers in that known direction and the best-scoring candidate becomes the student's training target. On facts about parents and their children, known-direction scoring yields more accurate labels than scoring the requested direction, even after tuned corrections for name priors. With prior-corrected scores, the better direction depends on the facts rather than the template, and reverses on mined facts whose notable entity is the parent rather than the child. With the evaluated children's forward facts withheld, students trained on known-direction labels improve open-ended accuracy on their trained queries by 13 to 15 points over students trained on prior-corrected reverse labels. After generated answers are matched to a fixed name list by lexical similarity, students reproduce nearly all selected labels. Their accuracy largely follows label quality. The label advantage holds on unscreened queries and when candidates are retrieved without inserting correct answers. Our findings show that directional verification mitigates the transfer of errors from teacher-generated answers to students by providing more accurate training targets. Code is available at https://github.com/js-lee-AI/directional-verification.

Understanding loss landscapes is central to explaining neural-network training, yet their structure remains only partially understood even in simple models. We study the Gaussian population loss of shallow, bias-free ReLU networks in the teacher--student setting. This provides a simple model for studying essential aspects such as feature learning and overparameterization. For teacher networks with positive output weights and planar features, we show that including a learned linear skip removes all spurious local minima with non-negative student output weights once the student network is at least as wide as the teacher network. In contrast, without the skip, we construct a fixed teacher network with positive output weights and only three hidden neurons in input dimension two whose spurious local minima persist at every student width at least three. Thus, a learned linear skip can remove spurious minima that persist under arbitrary overparameterization. Furthermore, we show that a positive output weight student network always learns the subspace spanned by the teacher features: student features at local minima with non-negative student output weights lie in the span of the teacher features. For ReLU networks in two dimensions, even heavily overparameterized student networks have effective width controlled by the teacher width: every critical point with positive student output weights has at most twice as many distinct student feature directions as teacher neurons. Finally, we transfer the benignity result to empirical minima over parameter balls of any prescribed radius, with the required sampling accuracy depending on that radius.

Recent advances in deep learning for shadow removal have significantly enhanced image quality and realism. However, most approaches rely on real-world paired datasets, which are costly to collect and often limited in scene diversity, leading to limited generalization. To address these limitations, we propose One-step Deshadow Diffusion via Reward guidance (ODDR), a new framework that achieves efficient and high-fidelity shadow removal without relying on real-world paired supervision. Our method begins with One-step Deshadow Diffusion (ODD), a baseline model trained on synthetic shadow data for efficient one-step shadow-free reconstruction. We further adapt ODD into ODDR using ShadowReward. In contrast to traditional, annotation-heavy approaches, ShadowReward is the first reward model for shadow removal trained entirely without human annotation. It learns to mimic human perceptual judgments by ranking synthetically generated images with controlled degradations, such as texture distortion and boundary artifacts. This reward-guided fine-tuning enables ODDR to close the synthetic-to-real domain gap. Extensive experiments show that ODD achieves strong performance without relying on real-world paired supervision, and ODDR further improves the results, narrowing the gap to fully supervised methods trained on real-world paired data while maintaining higher computational efficiency as a single-step model.

Active vision manipulation requires a policy to control both its camera and its end-effectors, yet camera motion determines which evidence remains visible within finite observation windows. Acquiring a new view can displace task-critical cues, while retaining a view forgoes potentially useful observations. We formulate this as an evidence-aware retain--acquire problem and present ActiveWAM, a unified world--action model that learns observation and manipulation jointly. To this end, we propose training-time inversion which constrains a frozen video prior by task-bearing source evidence and visible temporal changes, eliminating the need for test-time inversion or candidate ranking. At deployment, the policy generates bimanual and pan/tilt actions-including stay and reacquisition behaviors-from view-aware history, and updates context from newly measured RGB observations. Future-video prediction serves as a co-training signal, while action generation requires neither future-video decoding nor optimal viewpoint annotations. We introduce RoboTwin-AV, a 50-task benchmark with executable pan/tilt control and automatically generated demonstrations. ActiveWAM improves TAVIS out-of-distribution success by up to 17.0 percentage points over the strongest baselines, achieves 20.0 additional points over Fast-WAM on RoboTwin-AV, and outperforms it by 26.7 points on real-world physical kitchen tasks.

All the objects composing our world are enclosed within surfaces. Yet, most robot learning and motion generation frameworks treat surfaces as constraints ignoring their intrinsic geometry. This gap is acute for polyhedral meshes--the standard output of CAD and 3D reconstruction--whose discrete geometric structure remains unexploited. In this paper, we propose a unified discrete Riemannian framework that enables robot learning directly on polyhedral surface meshes. Using discrete differential geometry, we define logarithmic and exponential maps, parallel transport, and ambient-space projections that remain well-defined across faces, edges, and vertices. We instantiate the framework in three learning paradigms: (i) Dynamic Movement Primitives (DMPs), an improved exponential-map computation and a fixed-tangent-cone forcing-term encoding with parallel transport yield better cross-surface generalisation and stability over prior mesh-based approaches. (ii) Gaussian Process (GP), a geodesic-based kernel with practical admissibility control, enables regression at arbitrary mesh locations without smoothness assumptions. (iii) Riemannian Flow Matching (RFM), mesh-native operators improve generative quality over spectral baselines while reducing training time. The framework is validated in simulation against state-of-the-art methods and demonstrated on two real-robot scenarios: generalising user-drawn trajectories across different surfaces and planning polishing motions on RGB-D-reconstructed surfaces.