Papers

Reinforcement learning with verifiable rewards (RLVR) relies on costly model-generated responses, motivating larger rollout batches and square-root learning-rate scaling. Reported to approximately preserve progress per sample with one optimizer step per rollout, this rule is less understood in practical multi-step training. We develop a two-clock framework that separates the samples needed to make progress from those consumed before collapse. Under a local quadratic model, the square-root rule preserves first-order drift per sample but not progress: larger batches take fewer, longer updates and lose more to curvature, yielding a learning-rate-dependent critical batch beyond which sample efficiency degrades. A drift-threshold hypothesis predicts a complementary failure clock whose sample count remains approximately invariant along the square-root rule. Controlled GRPO experiments on Qwen2.5-Math-1.5B across a $16\times$ batch range support both predictions. Ratio clipping delays failure, especially for larger batches, but the progress gap persists without it. Widening the clip trades stability for a transient speed gain. The progress gap reappears on another dataset, a general-purpose checkpoint, and a 7B model. Finally, we use the two-clock framework to derive a simple multi-step RLVR schedule: start below the critical batch and switch to a larger batch before failure. Our code is available at: https://github.com/WinstonLiyt/two-clocks-rlvr.

Multi-objective molecular optimization is a critical process in drug discovery that aims to modify a given molecule to simultaneously improve multiple properties while maintaining a similar overall structure. This process often requires iterative modifications guided by computational or experimental feedback. Recent large language model (LLM)-based methods have begun to explore this iterative approach; however, these methods are not explicitly trained to leverage such feedback, while existing approaches primarily focus on single-turn molecular optimization. To address this gap, we introduce Multi-turn Feedback Guided (MFG) molecular optimization, a reinforcement learning framework that trains an LLM to progressively optimize molecules using feedback from previous turns. MFG introduces a feedback vector capturing both property improvements and source molecule similarity, alongside a feedback-aligned reward function that promotes progressive improvement across turns. We apply MFG over $\text{GeLLM}^{3}\text{O}$, a state-of-the-art LLM model for multi-objective molecular optimization, evaluate $\text{GeLLM}^{3}\text{O-MFG}$ on ten multi-objective molecular optimization tasks. The results demonstrate that $\text{GeLLM}^{3}\text{O-MFG}$ achieves the highest average similarity-constrained success rate ($81.50%$), outperforming the base model $\text{GeLLM}^{3}\text{O}$ ($70.81%$).

Robots are increasingly autonomous in task execution, yet their improvement remains largely human-driven. This study investigates Physical Recursive Self-Improvement (Physical RSI), in which physical experience guides lasting updates to both a robot's capabilities and its improvement procedure, making subsequent improvement more effective across tasks. Physical RSI faces challenges from noisy observations, continuous dynamics, costly trials, and irreversible consequences. We present an L0-L5 framework that differentiates capability-level from meta-level updates and within-task from across-task gains, reserving L5 for independently verified recursive gain. We study existing methods around harness adaptation and model training, alongside safety and control mechanisms that govern execution and successive updates. Our analysis reveals that, despite substantial progress in self-improvement, the surveyed literature does not yet demonstrate Physical RSI as defined by our framework. We identify research priorities in reliable feedback, causal attribution, safety preservation, and benchmarking, outlining a path toward robots that autonomously improve how they learn while ensuring safety. The GitHub repository for Physical RSI is available at https://github.com/SAIL-Research-Lab/Physical-RSI.

This paper investigates offline Q-learning for finite-horizon Markov decision processes with Kullback-Leibler (KL) regularization. Recent studies have shown that KL regularization can yield fast rates in offline decision-making problems. However, it remains unclear whether Q-learning, which adopts incremental stochastic approximation updates, can achieve such fast rates in the offline setting. We first propose KL-regularized Q-Learning (KLQ), a simple variant of classical Q-learning. By exploiting KL regularization, we prove that KLQ achieves a sample complexity of $\widetilde{O}(1/\epsilon)$ without explicit pessimism. This guarantee, however, has an exponential dependence on the KL regularization parameter and the horizon. To address this limitation, we further propose KL-regularized Pessimistic Q-Learning (KLPQ), which incorporates pessimism into KLQ. We prove that KLPQ retains the same \(\widetilde O(1/\epsilon)\) rate while removing the explicit exponential dependence, thereby further improving the sample complexity. To the best of our knowledge, these results establish the first fast-rate guarantees for offline Q-learning in tabular finite-horizon KL-regularized MDPs.

A central challenge in scaling flow models is maintaining stability in training and integration. In flow matching, this stability depends critically on how source samples are paired with data. Independent random pairing introduces substantial stochasticity, motivating minibatch optimal transport (OT) to construct online source-target couplings that reduce this randomness (Tong et al., 2024). However, as applications extend to larger and higher-dimensional datasets, small minibatches can limit these benefits, while scaling the coupling to substantially larger minibatches at every iteration is computationally prohibitive. Moreover, how OT coupling affects the full stochastic gradient and integration remains theoretically unresolved. In this work, we develop a scalable online entropic OT approximation that enables substantially larger source-target couplings while preserving the benefits of OT. We close the theoretical gap for both unregularized and entropic OT by proving strict reductions in full stochastic-gradient variance and integration error relative to independent pairing, and further characterize how these improvements strengthen as the online coupling minibatch grows. Experiments on single-cell gene interpolation-flow and image-flow training validate these predictions, showing that larger online couplings consistently improve training stability and final generation performance.

It has been difficult to accurately model the physics of the interaction between a robot and the objects it manipulates, including contact mechanics, coupled contact dynamics, deformations, stiffness and friction. We present RoboSimStudio (RSS), an interactive toolkit that makes choices of solver, contact model, and physical parameters explicit and editable, allowing both humans and coding agents to quickly create and iteratively refine simulations based on observed physical behavior. We implement RSS with the open Newton physics simulation system and describe experiments with human-guided real2sim using (1) visual evaluation, (2) force-feedback and (3) agentic Real2Sim2Real. Code, video and additional information can be found on the project website: https://robosimstudio.github.io

Simulation-based Bayesian inference for Chemical Reaction Networks (CRNs) must simultaneously solve two computationally demanding problems: discriminating between overlapping reaction pathways and estimating reaction rate constants. Classical approaches, such as Sequential Monte Carlo ABC, rely on handcrafted summary statistics and scale poorly across large parameter spaces. To overcome this, we introduce a neural inference architecture. Our pipeline utilizes shared 1D-CNNs and a permutation-invariant Set Attention mechanism to learn low-dimensional summary statistics from variable-length, multi-experiment time-series data. These statistics simultaneously condition a model selection network to compute posterior model probabilities and model-specific conditional Invertible Neural Networks (cINNs) that estimate parameter posterior distributions. We demonstrate the practical utility of our approach on a real protodeboronation example, illustrating how it handles structural ambiguity and extracts kinetic parameters from competing catalyst deactivation pathways.

In this work, we study grokking in a simple, analytically tractable setting whose rich phenomenology allows us to disentangle the effects of initialization, signal strength, weight decay, sample size, and input dimension. These factors can produce superficially similar learning curves through distinct mechanisms, complicating their interpretation. Our analysis distinguishes early fitting through nuisance features, core feature learning, and delayed generalization. We identify sharp conditions under which grokking provably occurs and derive analytic asymptotic expressions for the lengths of the generalization plateaus. Our results also characterize regimes of early generalization and failure to reach a prescribed test accuracy target. We show that the interpolation edge constitutes a distinct dynamical regime, where vanishing spectral gaps separate core learning from generalization, helping clarify the role of proximity to interpolation in existing accounts of grokking. The resulting picture is analogous to fast and slow phases in chemical kinetics and suggests that the lazy-to-rich transition discussed in the literature may reflect a similar separation of dynamical time scales. Our theoretical results are faithfully supported by extensive experiments.

Residual connections propagate information through a network by additively accumulating layer outputs. Recently proposed alternatives, such as Manifold-Constrained Hyper-Connections (mHC) and Attention Residuals (AttnRes), instead use more expressive depth aggregation mechanisms that improve language modeling performance at scale. In this work, we consider the mixing of representations over depth as a problem of storing and retrieving information across layers. Formalizing this perspective, we introduce Depth Memory (DM), a unified framework that describes the standard pre-layer-norm (PreLN) transformer, mHC, and AttnRes in terms of streams that carry representations through the network. One such stream, for example, could be an explicit memory stream that is queried to enrich the input to every layer, and whose output is subsequently written back into the memory for use by later layers. For AttnRes, this memory stream is a growing cache of explicit layer representations. For mHC, it is an expanded residual stream. Building on this view, we propose Linear Depth Memory (LDM), which formulates information retrieval across depth as a regression problem solved using recursive least squares. The resulting architecture propagates three streams across depth: the standard residual stream, a linear associative memory stream, and an inverse covariance stream. For high-throughput training and inference, we propose a hardware-aware implementation of LDM that reduces peak activation memory and memory traffic. LDM consistently achieves the lowest perplexities among PreLN, mHC and AttnRes baselines at the 400M and 1.5B parameter scales, alongside competitive zero-shot accuracy.

Agentic reinforcement learning (RL) enables large language models (LLMs) to learn from multi-turn interactions with verifiable outcomes. Textual skills can improve exploration by distilling prior interactions into reusable guidance. However, existing skill-based RL approaches often either assume skill injection is beneficial or rely on indirect feedback or rollout-intensive evaluation to assess its utility. Our pilot studies show that rollout-based utility estimation for skill credit assignment is costly and unstable under limited sampling, while skill benefits vary across candidates and depend on the current policy's competence on each query. To address these challenges, we introduce SkillScaffold, a policy-adaptive training scaffold that combines low-cost skill-utility estimation with selective skill injection. By reusing outcome-verified training trajectories, SkillScaffold generates candidate skills and estimates their utility through outcome-aware likelihood comparisons, without candidate-specific environment rollouts. The resulting utility guides whether and which skill to inject, retaining skill-free execution when no candidate is estimated to help. Joint training enables the Generator to produce skills tailored to the evolving Executor. Scaffolded RL encourages skill internalization for skill-free inference. Experiments on ALFWorld, WebShop, ScienceWorld, and $\tau^2$-Bench demonstrate gains over strong agentic-RL and skill-based RL baselines across both short- and long-horizon tasks. Ablations and further analyses support the importance of utility-based selection and abstention. Our code is available at https://github.com/Vaidurya00/SkillScaffold.

Raw image intensities are a poor surrogate for similarity across images in correspondence problems. Further, while auxiliary segmentation labels can provide semantic identification and improve matching, they can assign only coarse identities to fine-grained anatomy and often cover only a subset of structures, particularly in whole-body settings. We present an approach that instead establishes correspondence using complementary representations of anatomy and demonstrate its utility for longitudinal whole-body PET/CT registration in the Learn2Reg 2026 PSMAReg Challenge. Our 3D foundation model, anatomix, is used to derive fine-grained, per-voxel semantic representations that are consistent across volumes, and we further employ off-the-shelf segmentation models to obtain coarse anatomical identities. Together, these representations drive a GPU-accelerated iterative registration solver (FireANTs) in a hybrid multistage framework: the first stage estimates alignment from dense multi-channel features, after which a second refines the deformation using both anatomix- and segmentation-based losses. Lastly, because registration should not introduce non-biological tumor volume changes across time, we develop lesion-specific component-wise volume preservation schemes. Without any task-specific training, our system won first place in the PSMAReg Challenge leaderboard. Code and interactive visualizations are available at http://www.neeldey.com/reps-for-reg.

Gauge-equivariant graph neural networks generate ordered products of gauge links through local message passing, but the geometry of these internal structures is not captured by the receptive field alone. We investigate the statistical ensemble of closed lattice walks generated by recursive node- and edge-feature operations, independently of network training. Analytical and numerical results reveal a separation between rapid growth of contour length and much slower spatial expansion: deeper compositions generate increasingly long paths through repeated traversal of relatively compact regions. Loop expansions of fermionic observables provide a physical interpretation of this hierarchy, showing how spatial locality remains compatible with high-order dependence on nearby gauge fields. These findings identify complementary roles of network depth in extending spatial reach and enriching the representations constructed within that reach, providing a geometrical perspective on learning lattice gauge observables.

We develop a rate-transfer method that converts finite-time Lyapunov estimates for an associated stochastic differential equation (SDE) into finite-time convergence bounds for stochastic gradient algorithms. Under generic moment and smoothness conditions, the convergence analysis of the stochastic gradient method reduces to a calculus argument verifying a Lyapunov generator inequality for the SDE. The method recovers established rates for classical stochastic gradient methods and provides a way to establish convergence rates for new optimization algorithms through calculus arguments applied to their associated SDEs.

Off-policy data in reinforcement learning with verifiable rewards (RLVR) can arise from asynchronous generation, rollout-training mismatch, and response reuse. Token-wise importance weighting avoids multiplying likelihood ratios across a response but generally biases policy gradients. We prove that, under coverage and local expressiveness conditions, full future KL credit makes the exact proximal optimizer the update's unique stationary policy, even for mixtures of older policies, without full-response weights or a fitted critic. We further establish convergence of continuous-time tabular population updates under persistent coverage and time-varying behavior policies, and a finite-sample reward guarantee for linear logistic models under additional assumptions. These results motivate Proximal KL, a practical off-policy RLVR method that separates the proximal reference from the sampling policy and replaces PPO clipping with bounded future KL credit. Across all seven matched off-policy comparisons on Qwen3-1.7B and Qwen3-4B, spanning two learning rates, response reuse, and rollout staleness, Proximal KL outperforms GRPO in micro-averaged accuracy over eight math benchmarks by up to 2.58%. A separate on-policy comparison yields comparable aggregate accuracy. In periodic-refresh experiments, Proximal KL avoids the collapses observed in PG and CISPO, exceeds GRPO's training accuracy, and reaches higher accuracy in fewer policy updates. Ablations show that token-wise correction outperforms sequence-level and unweighted controls. With token-wise correction retained, future KL feedback reduces sensitivity to the tested baseline perturbation relative to PPO clipping. A finite future horizon improves AIME24 mean@16 and pass@16 over the full tail, while fixed-policy measurements show that truncation reduces correction variance over the tested horizons.

Speculative Decoding (SD) has demonstrated its efficacy in accelerating large language model inference, but for Vision-Language Models (VLMs), the diverse input modalities make it difficult for a small trained drafter to achieve consistently high throughput. Self-speculative decoding reuses parts of the target model as a drafter which retains the multi-modal capability of the VLMs, yet its efficiency depends on choosing a suitable draft configuration for each prompt. We propose Single-Peak SPeculative decoding (SP$^2$ec), a training-free bandit algorithm that adapts the draft configuration during generation. SP$^2$ec exploits the single-peak throughput pattern observed in existing self-speculative methods, including SWIFT, KnapSpec, SpecVLM, and our simple cosSim. Our algorithmic design explicitly minimizes a novel notion of wall-time regret for each given query, corresponding to the empirical wall-time latency. Theoretical guarantees on the wall-time regret show that SP$^2$ec can exploit the optimal draft configuration, with negligible exploration cost. Extensive experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-Mistral-7B across visual and textual tasks show consistent improvements of up to 11% over the original self-speculative methods, demonstrating the wide applicability of the proposed SP$^2$ec algorithm.

Manufacturers often need to characterize a new material, process, or product before enough operating data are available to train reliable predictive models for this new, target domain. Existing manufacturing operations contain potentially useful knowledge to accelerate this characterization, but indiscriminate reuse of data from existing source domains can cause negative transfer in learning of the new, target domain when the manufacturing domains differ sufficiently. This work proposes a reinforcement-learning method for selective knowledge transfer, where its reward scores each opportunity to transfer against supervised alternatives, so negative transfer is avoided and controlled during forward prediction instead of being diagnosed afterwards. Given limited data from a new target domain, the learned policy chooses among predictors based on individual source operations, subsets of the sources, or the supervised alternatives: fully pooled sources, or target data alone. A validation gate deploys the transferred prediction only when it improves upon the supervised alternatives. The method is designed for learning variables supporting carbon accounting, as well as performance and quality outcomes. We demonstrate the approach by quantifying the embodied greenhouse-gas emissions of parts produced via material-extrusion additive manufacturing, using data from 729 experiments spanning five materials: ABS, PETG, TPU, PLA, and recycled PLA. The response combines measured fabrication electricity with polymer-specific cradle-to-gate production factors, while 22 material, process, and toolpath features characterize each experiment. In leave-one-material-out evaluation, this method achieves a mean test RMSE of 5.2 g CO2e per part, 20.00% lower than target-only training (32.5% lower when only 5-10% of the target training data are available) and 10.30% lower than indiscriminate source-target pooling.

Natural Language Autoencoders (NLAs) are proposed as oversight tools for language models: an Activation Verbaliser (AV) writes an English explanation of an activation vector, and an Activation Reconstructor (AR) maps the explanation back to a vector. Their value for oversight depends on a human monitor being able to tell a true account of an activation from a false one, given the Av's explanation. We formalise NLA training and NLA oversight deployment as two different referential games from Emergent Communication: training is a generative referential game, while oversight is a discriminative, multi-modal, partially-observable referential game played over a many-to-many relation between the monitored language model's activations and the accounts that are true of them. The gap between the two games predicts that AV's explanations will identify which activation they describe without discriminating its true account from closely matched false ones, since training never asks explanations to rule anything out. Using the NLA's own AR as a maximally-sympathetic surrogate monitor model, we experimentally confirm this theoretical insight on one released checkpoint: true accounts beat accounts of other activations in all 320 comparisons, but beat fluent rivals differing in a single claim only 60.2% of the time (std.err. 6.7%, chance at 50%, $n = 35$). A meaning-preserving rewrite by an uninvolved model costs only 2.1% of the distance to another activation's account, so the failure is not due to the NLA pair having developed private conventions in the AV's explanations. Our rival accounts test only transcript-level truths, which the monitored language model's conversation settles, so these results bound the easiest case and leave claims about the monitored model's own processing untested. We state conjectures that would make both results consequences of the framework.

A transformer trained on sequences whose source parameters are drawn afresh each time can learn only what the sequences share: the source family and its prior. The Bayesian predictor that knows both, which we call informed Bayes, is the unique minimizer of the training loss and is computable from the family alone; we use it as a quantitative, falsifiable reference. Where earlier work compared transformers with Bayes on sources whose contexts can be learned independently, we build Markov sources with up to 64 symbols in which observations in one context inform prediction in another: product sources with known or hidden encodings, transition rows sharing an unknown base law and depth, and a mixture of three families. Our main result is that informed Bayes predicts, from the family and prior alone, how a trained transformer behaves in four tests that go beyond the learning curve: the loss that remains under a source outside the training family (predicted 2.237 bits, measured 2.309), the doubling of the extra loss when two independent components change rather than one (0.345 predicted, 0.350 measured), the benefit of other contexts when a token appears for the first time, and the depth at which copying an earlier continuation starts to pay. On the learning curves themselves, small transformers match informed Bayes when the shared structure is visible in the tokens, and hidden structure needs more layers and training, with gaps below 0.01 bits per token for the largest models. We explain why the imitation is possible and where it must fail: informed Bayes decomposes into sufficient statistics that attention selects and accumulates and an estimator, fixed by the prior, that the feed-forward network learns; training pays the complexity of informed Bayes under the architecture's prior, and finite-memory universal prediction bounds how closely a bounded network can reach it.

The rapid advancement of image-to-text extraction systems and large language models (LLMs) poses new challenges for educational assessment by making assignment materials easy to transcribe and solve with minimal human effort. To address this problem, we present AIRRI (Artificial Intelligence Resilient Reading Interface), a framework for generating machine-resistant educational content, a dataset of higher-education assignments for evaluation, and an initial evaluation of the human readability of AIRRI machine-resistant text designs. AIRRI employs two categories of machine-resistance techniques. First, AIRRI applies image transformations including character spacing, word spacing, line spacing, character opacity variation, stripe overlays, and image noise. Second, AIRRI applies decision-based adversarial attacks: RayS, SurFree, and ADBA. We evaluate AIRRI against three image-to-text systems: Tesseract, EasyOCR, and the transformer-based GOT-OCR2 model. Our results show that AIRRI substantially degrades automated text extraction performance. AIRRI increased character error rates for EasyOCR (6.9% to 55.9%), Tesseract (2.3% to 100.0%), and GOT-OCR2 (30.7% to 48.1%). Word error rates increased for EasyOCR (24.9% to 96.7%) and Tesseract (10.7% to 100.0%). Among the evaluated decision-based attacks, ADBA consistently produced the largest degradation in image-to-text performance, while RayS and SurFree produced more moderate reductions. EasyOCR and Tesseract were generally more susceptible to both image-based and adversarial perturbations than the transformer-based GOT-OCR2 model. Our findings suggest that AIRRI can significantly reduce the reliability of image-to-text-to-LLM workflows, increasing the effort required for AI-assisted assignment completion.

Non-clashing teaching minimizes teaching complexity among models satisfying the Goldman–Mathias collusion-avoidance criterion. While the Vapnik–Chervonenkis dimension (VCdim) measures the complexity of learning from samples, non-clashing teaching dimension (NCTD) measures the number of examples required when a teacher can select them. Whether NCTD is always bounded above by VCdim remains a fundamental open question in learning theory. In this paper, we investigate this question for the concept classes of closed neighborhoods in planar graphs. We prove a tight upper bound of two on their NCTD, thereby establishing the conjectured inequality for this family. For the positive variant NCTD+, in which teaching sets contain only positive examples, we obtain a tight upper bound of four. Finally, we show that deciding whether either NCTD+ or NCTD is at most one is NP-complete, even for concept classes of closed neighborhoods in planar graphs. In particular, despite the upper bound of two on NCTD for such classes, distinguishing between NCTD values of one and two remains NP-hard.

World models (WMs) and world-action models (WAMs) focus on future world prediction and executable action generation, respectively. Although these models learn shared physical dynamics and interaction semantics from embodied trajectories, they are typically trained in isolation, which restricts knowledge reuse between prediction and control. The central challenge is therefore to develop a unified learning framework that shares these dynamics while preserving the distinct modeling capabilities required by each paradigm. We propose UniW(A)M, a mixture-of-experts world-(action) model built upon diffusion Transformers that unifies action-conditioned world modeling and diverse WAM variants. Shared experts capture common dynamics, while atomic capability experts are reused across interfaces so that each interface keeps its own conditional pathway. At deployment, a plug-in adaptive interface router (PAIR) learns from a small subset of target-task chunk scenarios to select a suitable WAM interface for each execution chunk, balancing task success and execution efficiency while keeping the backbone frozen. We evaluate UniW(A)M on LIBERO, RoboTwin 2.0, and three real-world manipulation tasks, where it achieves 88.3% average real-world success versus 76.7% for the strongest baseline. On WorldArena, it leads action-conditioned world modeling on all three tasks. AC-WM co-training contributes to action generation on every action interface and PAIR raises OOD success from 63.3% to 76.7%.

Kronecker-preconditioned optimizers recompute a cached curvature eigendecomposition only periodically, and implementations and adaptive methods control this through the refresh frequency. We show that frequency leaves part of a schedule free: at a fixed number of refreshes, when they occur changes held-out loss at the same compute. Training GPT-style models with the Mousse optimizer across model sizes, we move the middle refreshes of a four-refresh schedule while fixing the first and last. On FineWeb the two placements exchanged rank as the learning rate rose (later placement was better at low rates, earlier placement at high ones), so a single-learning-rate comparison can report either ranking. The reversal held under controls including the published optimizer internals, and placement still mattered at the published refresh density. The learning rate at which the preference flips fell with model size, consistent with a held-out smaller model, and shifted with training length. A spectral-width staleness score predicted the same ranking at every learning rate and missed the reversal. Refresh schedules should be reported in full and compared across the learning rates at which they will be used.

Vision-language models (VLMs) struggle with metric and multi-view relationships that remain implicit in RGB observations. Geometry foundation models recover this structure from RGB, but existing methods either fuse the estimated features with visual tokens and learn to use them from question answering alone, or align VLM latents with geometry features. Capturing geometry, however, does not ensure its use: zeroing the aligned latents of a released feature-alignment model leaves its accuracy unchanged. We introduce GeoReader, a framework that learns geometric representations for the VLM that reads them. First, we train lightweight Geometric Readers to produce continuous input tokens, supervising them with geometry descriptions decoded by a frozen VLM, so that the Readers express geometry in a form the target model can already read. We then freeze the Readers to preserve this interface while fine-tuning the VLM on spatial QA, with geometry injection conditioned on the question. Geometry descriptions serve only as training targets: at inference, the VLM answers directly from the continuous geometric tokens without generating a geometry chain of thought. With geometry predicted from RGB images, our approach reaches 94.3 on MindCube-Tiny with Qwen3-VL-4B and, with a multi-source QA mixture, an average of 72.5 across six additional spatial benchmarks with Qwen3.5-9B. These results suggest that a useful geometric representation is one the receiving model can use, not merely one from which geometry can be reconstructed.

Image-token compression must reduce computation without hurting the semantics. We introduce Varest, guided by semantic fungibility: tokens within a coherent semantic region can be summarized coarsely, while object boundaries and literal text require finer distinctions. Object and text localizers guide variable-resolution merging, preserving spatial coverage and finer detail at boundaries and text. Learned [REL] registers then prune visual tokens after early decoder processing. Training across budgets enables one checkpoint per backbone. Across three backbones and 13 benchmarks, Varest outperforms 15 baselines in mean performance retention at every evaluated budget, retaining 86.10-92.33% of full-token performance with only 5.56% of the original layer-averaged image-token count.

Personal-agent memory has largely focused on learning what users know, prefer, and want. However, knowing the user does not necessarily mean knowing what the user and the agent have established together. An assistant's proposal should not become an authorized instruction merely because it is remembered. We introduce Grounding Awareness, the ability to distinguish accepted, unacknowledged, revised, and scoped arrangements when selecting actions. We develop a controlled benchmark to test whether agents can tell the difference between proposals users have accepted and those they have not, and whether they apply accepted arrangements only where appropriate. Across multiple agents and memory systems, we find that strong memory performance does not ensure Grounding Awareness. More surprisingly, providing more complete user information can increase execution of unacknowledged proposals, while persistent memory can cause agents to treat these proposals as if they were user instructions. We trace these failures through memory writing, retrieval, and action selection. Finally, we introduce Grounding Layer, a lightweight, training-free extension that explicitly tracks grounding status and checks consequential actions, substantially improving agent behavior without modifying models or memory systems.

Reinforcement learning (RL) has emerged as the primary paradigm for optimizing Large Language Model (LLM) agents, yet efficient critic-free algorithms (e.g., GRPO) suffer from a fundamental uniform credit assignment dilemma that broadcasts trajectory-level advantages equally to all actions. Our analysis reveals that this uniformity inherently reinforces structurally invalid actions in preferred trajectories and ignores the downstream cascading influence of early pivotal deviations in unpreferred ones. To resolve this without introducing critic overhead, we propose Progress-Aware Advantage Shaping (ProAS), a plug-and-play framework that leverages a dynamically evolving Agentic Reference Graph (ARG) to achieve fine-grained credit assignment via Masking Unreasonable Incentives (MUI) and Asymmetric Penalty Redistribution (APR). Extensive experiments across Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct demonstrate that ProAS universally elevates mainstream group-based baselines to new state-of-the-art performance, achieving peak success rates of 91.9% and 94.3% on ALFWorld, alongside 66.7% and 70.8% on WebShop, respectively. Furthermore, it accelerates training convergence by up to $2.44\times$ while maintaining strong robustness against imperfect or distilled ARGs. Codes: https://github.com/LINs-lab/ProAS.

Online decision-making faces a tension between efficient inference and long-horizon reasoning. Value-based methods rely on short-horizon bootstrapping that limits the propagation of long-term information, while trajectory optimization methods plan explicitly but must solve a high-dimensional optimization problem at every decision step. We propose Generative Trajectory Planning (GTP), a model-based framework that addresses this tension by learning a generative model as a proposal distribution over action sequences. Instead of optimizing sequences from scratch, GTP samples candidate sequences from a diffusion prior and refines them using learned dynamics, reward, and value models, enabling efficient trajectory-level reasoning. Across continuous control and long-horizon manipulation tasks, GTP matches or exceeds state-of-the-art model-free, diffusion-policy, and model-based baselines, and remains stable in regimes where strong baselines degrade. Our results suggest that generative models are best understood not as policies, but as structured proposal distributions for planning, offering a new perspective on how to integrate generative modeling with model-based decision-making.

Generative models are often required to serve many downstream conditional tasks. In the training-free setting, a pretrained unconditional model is reused as a prior and combined with a task-specific likelihood to sample from the corresponding posterior. While such training-free samplers have been developed for diffusion and flow models, their counterpart for Schrödinger bridge (SB) models remains unexplored, despite the growing appeal of SBs as flexible stochastic transport models between general source and target distributions. We propose BridgeTwist, a training-free conditional sampler for pretrained SB generative models, formulated as a tempered twisted sequential Monte Carlo scheme. Its twist is a closed-form plug-in surrogate, obtained by combining the velocity and score fields exposed by SB pretraining into an endpoint predictor and evaluating the likelihood there. Since this surrogate is less reliable near the source, we temper the twist with a time-varying exponent, downweighting its influence at early steps. We further introduce a bridge-shaped inference-time diffusion schedule that concentrates the proposal noise in the interior of the path. Telescoping importance weights then cancel all intermediate twist factors, yielding an asymptotically exact sampler for the posterior. Empirically, BridgeTwist outperforms training-free baselines on class-conditional and text-prompted image generation, and is competitive with specialized inverse problem solvers on inpainting, deblurring, and super-resolution.

Language model systems repeatedly make test-time decisions by comparing similarity: they retrieve exemplars, rank candidates, and revise responses. Yet similarity is not unique: lexical, semantic, structural, and stylistic resemblance can point to different candidates, and the task-aligned notion can change across instances. We introduce BASIS (Bayesian Axis Selection via Interactive Signals), a training-free Bayesian method that uses weak test-time feedback to infer which criterion in a supplied, task-specific library governs the current decision. It maintains a joint posterior over candidate targets and criteria, updates it from scalar scores, pairwise preferences, or verifier outputs, and selects probes by expected information gain. In a finite noisy model, we characterize when the target-criterion pair is recoverable and show why fixed-criterion and fixed-mixture rules can fail under criterion shift. Across concept recovery, reasoning exemplar selection, and response refinement, BASIS improves six-probe concept-recovery Success@1 from 56.0% to 68.1%, reasoning accuracy from 72.3% to 73.1%, and refinement win rate from 61.4% to 63.8% over tuned static mixtures. It also exceeds compute-matched best-of-N selection on reasoning (72.1%) and refinement (61.2%). Gains are largest under criterion mismatch and with larger candidate pools, and matched controls indicate that adaptive criterion weighting explains much of the benefit, with joint target-criterion inference adding further value. BASIS selects only among supplied criteria and does not discover new ones.

Texture interpolation asks what should lie between two textures. Standard methods typically blend pixels or features at corresponding spatial locations, implicitly assuming that co-located elements should be mixed. For unrelated textures, however, this correspondence is arbitrary: a crack at one location has no particular relationship to a fiber at the same location in another image. We study whether this hidden pairing choice is itself a major determinant of interpolation quality. We represent textures as collections of local appearance vectors, or textons, and replace positional pairing with appearance-aware matching while keeping the encoder, representation, and decoder fixed. This isolates the effect of correspondence from changes in model capacity or training. Across texture pairs, appearance-aware matching consistently improves the distributional quality of interpolated midpoints over positional feature blending. Soft matching produces the strongest midpoint statistics, whereas one-to-one assignments yield more balanced transitions and closer agreement with the target endpoint. The same intervention also improves a second pretrained texture interpolator without retraining. These results identify correspondence as a first-class design choice in texture interpolation: meaningful transitions depend not only on how strongly two textures are mixed, but also on which local appearances are mixed with one another.