In adversarial multi-armed bandits, two performance measures are commonly used: static regret, which compares the learner to the best fixed arm, and dynamic regret, which compares it to the best sequence of arms. While optimal algorithms are known for each measure individually, there is no known algorithm achieving optimal bounds for both simultaneously. Marinov and Zimmert [2021] first showed that such simultaneous optimality is impossible against an adaptive adversary. Our work takes a first step to demonstrate its possibility against an oblivious adversary when losses are deterministic. First, we extend the impossibility result of Marinov and Zimmert [2021] to the case of deterministic losses. Then, we present an algorithm achieving optimal static and dynamic regret simultaneously against an oblivious adversary. Together, they reveal a fundamental separation between adaptive and oblivious adversaries when multiple regret benchmarks are considered simultaneously. It also provides new insight into the long open problem of simultaneously achieving optimal regret against switching benchmarks of different numbers of switches. Our algorithm uses negative static regret to compensate for the exploration overhead incurred when controlling dynamic regret, and leverages Blackwell approachability to jointly control both regrets. This yields a new model selection procedure for bandits that may be of independent interest.

Streaming recurrent models enable efficient 3D reconstruction by maintaining persistent state representations. However, they suffer from catastrophic forgetting over long sequences due to balancing historical information with new observations. Recent methods alleviate this by deriving adaptive signals from attention perspective, but they operate on single dimensions without considering temporal and spatial consistency. To this end, we propose a training-free framework termed TTSA3R that leverages both temporal state evolution and spatial observation quality for adaptive state updates in 3D reconstruction. In particular, we devise a Temporal Adaptive Update Module that regulates update magnitude by analyzing temporal state evolution patterns. Then, a Spatial Contextual Update Module is introduced to localize spatial regions that require updates through observation-state alignment and scene dynamics. These complementary signals are finally fused to determine the state updating strategies. Extensive experiments demonstrate the effectiveness of TTSA3R in diverse 3D tasks. Moreover, our method exhibits only 1.33x error increase compared to over 4x degradation in the baseline model on extended sequences of 3D reconstruction, significantly improving long-term reconstruction stability. Our codes are available at https://github.com/anonus2357/ttsa3r.

Large language models have advanced from single-turn question answering to deep research systems that iteratively decompose research questions, invoke retrieval tools, and synthesize information across multiple rounds. Evaluating such systems typically involves scoring their final research reports holistically, but this end-to-end paradigm tightly couples the language model's decision-making, workflow design, and environmental feedback, precluding decomposable analysis of individual components. We introduce ScholarGym, an evaluation environment that isolates the information-gathering stage of deep research on academic literature. Under a unified workflow, ScholarGym decomposes the research process into three explicit stages -- Query Planning, Tool Invocation, and Relevance Assessment -- and evaluates each against 2,536 expert-annotated queries over a static corpus of 570K papers with deterministic retrieval. Systematic experiments reveal that iterative query decomposition yields 2.9--3.3$\times$ F1 gains over single-query retrieval, models with extended thinking trade recall for precision, and Query Planning quality together with Relevance Assessment constitute dual bottlenecks that separate proprietary from open-source model performance.

Existing value-based online reinforcement learning (RL) algorithms suffer from slow policy exploitation due to ineffective exploration and delayed policy updates. To address these challenges, we propose an algorithm called Instant Retrospect Action (IRA). Specifically, we propose Q-Representation Discrepancy Evolution (RDE) to facilitate Q-network representation learning, enabling discriminative representations for neighboring state-action pairs. In addition, we adopt an explicit method to policy constraints by enabling Greedy Action Guidance (GAG). This is achieved through backtracking historical actions, which effectively enhances the policy update process. Our proposed method relies on providing the learning algorithm with accurate $k$-nearest-neighbor action value estimates and learning to design a fast-adaptable policy through policy constraints. We further propose the Instant Policy Update (IPU) mechanism, which enhances policy exploitation by systematically increasing the frequency of policy updates. We further discover that the early-stage training conservatism of the IRA method can alleviate the overestimation bias problem in value-based RL. Experimental results show that IRA can significantly improve the learning efficiency and final performance of online RL algorithms on eight MuJoCo continuous control tasks.The code is available at https://github.com/2706853499/IRA.

The current state of event detection research has two notable re-occurring limitations that we investigate in this study. First, the unidirectional nature of decoder-only LLMs presents a fundamental architectural bottleneck for natural language understanding tasks that depend on rich, bidirectional context. Second, we confront the conventional reliance on Micro-F1 scores in event detection literature, which systematically inflates performance by favoring majority classes. Instead, we focus on Macro-F1 as a more representative measure of a model's ability across the long-tail of event types. Our experiments demonstrate that models enhanced with sentence context achieve superior performance over canonical decoder-only baselines. Using Low-Rank Adaptation (LoRA) during finetuning provides a substantial boost in Macro-F1 scores in particular, especially for the decoder-only models, showing that LoRA can be an effective tool to enhance LLMs' performance on long-tailed event classes.

This note introduces Projected Microbatch Accumulation (PROMA), a reference-free proximal policy method that controls KL divergence by projecting away high-variance components of the policy gradient. Two variants are presented. In the accumulation-based variant, the running gradient is projected orthogonal to the sequence-wise log-probability gradients of each microbatch. In the intra-microbatch variant, a factored projection using dominant subspaces of activations and gradient outputs is applied independently within each microbatch, making it compatible with standard data-parallel training. Empirically, the accumulation variant achieves tighter per-step KL control than GRPO with PPO clipping, while the intra-microbatch variant achieves the best validation performance.

Existing adversarial attacks for VLP models are mostly sample-specific, resulting in substantial computational overhead when scaled to large datasets or new scenarios. To overcome this limitation, we propose Hierarchical Refinement Attack (HRA), a multimodal universal attack framework for VLP models. For the image modality, we refine the optimization path by leveraging a temporal hierarchy of historical and estimated future gradients to avoid local minima and stabilize universal perturbation learning. For the text modality, it hierarchically models textual importance by considering both intra- and inter-sentence contributions to identify globally influential words, which are then used as universal text perturbations. Extensive experiments across various downstream tasks, VLP models, and datasets, demonstrate the superior transferability of the proposed universal multimodal attacks.

Uniform-noise discrete diffusion and flow models (e.g., D3PM, SEDD, UDLM, DFM) generate sequences non-autoregressively by iteratively refining randomly initialized vocabulary tokens through multiple context-dependent replacements. These models are typically formulated as time-inhomogeneous CTMC/DTMC processes and sampled using independent Bernoulli change decisions at each discretization step. This induces Poisson-binomial variance in per-position jump counts that grows with the number of required edits, leading to the characteristic under-editing (residual noise) and over-editing (cascading substitutions) failure modes that degrade sample quality, especially under tight discretization budgets. In contrast, absorbing-state (mask-start) models avoid this instability by allowing each position to jump at most once. We propose Stratified Hazard Sampling (SHS), a training-free, drop-in, and hyperparameter-free inference principle for any sampler that admits a stay-vs.-replace decomposition. SHS models per-token edits as events driven by cumulative hazard (CTMC) or cumulative jump mass (DTMC) and places events by stratifying this cumulative quantity: with a single random phase per position, a token is updated whenever its accumulated hazard crosses unit-spaced thresholds. This preserves the expected number of jumps while achieving the minimum possible conditional variance among unbiased integer estimators (bounded by 1/4 for any fixed cumulative mass), without altering per-jump destination sampling and thus retaining multimodality. Experiments on uniform-noise discrete diffusion language models show that SHS consistently improves sample quality. We further show that SHS improves robustness under token-level blacklist filtering, with benefits increasing as lexical constraints grow more severe.

This paper proposes an epistemological shift in the analysis of large generative models, replacing the category ''Large Language Models'' (LLM) with that of ''Large Discourse Models'' (LDM), and then with that of Artificial Discursive Agent (ADA). The theoretical framework is based on an ontological triad distinguishing three regulatory instances: the apprehension of the phenomenal regularities of the referential world, the structuring of embodied cognition, and the structural-linguistic sedimentation of the utterance within a socio-historical context. LDMs, operating on the product of these three instances (the document), model the discursive projection of a portion of human experience reified by the learning corpus. The proposed program aims to replace the ''fascination/fear'' dichotomy with public trials and procedures that make the place, uses, and limits of artificial discursive agents in contemporary social space decipherable, situating this approach within a perspective of governance and co-regulation involving the State, industry, civil society, and academia.

Preference learning from human feedback has the ability to align generative models with the needs of end-users. Human feedback is costly and time-consuming to obtain, which creates demand for data-efficient query selection methods. This work presents a novel approach that leverages optimal experimental design to ask humans the most informative preference queries, from which we can elucidate the latent reward function modeling user preferences efficiently. We formulate the problem of preference query selection as the one that maximizes the information about the underlying latent preference model. We show that this problem has a convex optimization formulation, and introduce a statistically and computationally efficient algorithm ED-PBRL that is supported by theoretical guarantees and can efficiently construct structured queries such as images or text. We empirically present the proposed framework by personalizing a text-to-image generative model to user-specific styles, showing that it requires less preference queries compared to random query selection.

Memristor-based in-memory computing has emerged as a promising paradigm to overcome the constraints of the von Neumann bottleneck and the memory wall by enabling fully parallelisable and energy-efficient vector-matrix multiplications. We investigate the effect of nonlinear, memristor-driven weight updates on the convergence behaviour of neural networks trained with equilibrium propagation (EqProp). Six memristor models were characterised by their voltage-current hysteresis and integrated into the EBANA framework for evaluation on two benchmark classification tasks. EqProp can achieve robust convergence under nonlinear weight updates, provided that memristors exhibit a sufficiently wide resistance range of at least an order of magnitude.

We propose Terminal Velocity Matching (TVM), a generalization of flow matching that enables high-fidelity one- and few-step generative modeling. TVM models the transition between any two diffusion timesteps and regularizes its behavior at its terminal time rather than at the initial time. We prove that TVM provides an upper bound on the $2$-Wasserstein distance between data and model distributions when the model is Lipschitz continuous. However, since Diffusion Transformers lack this property, we introduce minimal architectural changes that achieve stable, single-stage training. To make TVM efficient in practice, we develop a fused attention kernel that supports backward passes on Jacobian-Vector Products, which scale well with transformer architectures. On ImageNet-256x256, TVM achieves 3.29 FID with a single function evaluation (NFE) and 1.99 FID with 4 NFEs. It similarly achieves 4.32 1-NFE FID and 2.94 4-NFE FID on ImageNet-512x512, representing state-of-the-art performance for one/few-step models from scratch.

Large Language Models (LLMs) are increasingly using external web content. However, much of this content is not easily digestible by LLMs due to LLM-unfriendly formats and limitations of context length. To address this issue, we propose a method for generating general-purpose, information-dense summaries that act as plain-text repositories of web content. Inspired by Hegel's dialectical method, our approach, denoted as Chain of Summaries (CoS), iteratively refines an initial summary (thesis) by identifying its limitations through questioning (antithesis), leading to a general-purpose summary (synthesis) that can satisfy current and anticipate future information needs. Experiments on the TriviaQA, TruthfulQA, and SQUAD datasets demonstrate that CoS outperforms zero-shot LLM baselines by up to 66\% and specialized summarization methods such as Chain of Density, BRIO and PEGASUS by up to 27\%. CoS-generated summaries yield higher Q\&A performance compared to the source content, while requiring substantially fewer tokens and being agnostic to the specific downstream LLM. CoS thus resembles an appealing option for website maintainers to make their content more accessible for LLMs, while retaining possibilities for human oversight.

Accurate population estimates are essential for wildlife management, providing critical insights into species abundance and distribution. Traditional survey methods, including visual aerial counts and GNSS telemetry tracking, are widely used to monitor muskox populations in Arctic regions. These approaches are resource intensive and constrained by logistical challenges. Advances in remote sensing, artificial intelligence, and high resolution aerial imagery offer promising alternatives for wildlife detection. Yet, the effectiveness of deep learning object detection models (ODMs) is often limited by small datasets, making it challenging to train robust ODMs for sparsely distributed species like muskoxen. This study investigates the integration of synthetic imagery (SI) to supplement limited training data and improve muskox detection in zero shot (ZS) and few-shot (FS) settings. We compared a baseline model trained on real imagery with 5 ZS and 5 FS models that incorporated progressively more SI in the training set. For the ZS models, where no real images were included in the training set, adding SI improved detection performance. As more SI were added, performance in precision, recall and F1 score increased, but eventually plateaued, suggesting diminishing returns when SI exceeded 100% of the baseline model training dataset. For FS models, combining real and SI led to better recall and slightly higher overall accuracy compared to using real images alone, though these improvements were not statistically significant. Our findings demonstrate the potential of SI to train accurate ODMs when data is scarce, offering important perspectives for wildlife monitoring by enabling rare or inaccessible species to be monitored and to increase monitoring frequency. This approach could be used to initiate ODMs without real data and refine it as real images are acquired over time.

Instruction-guided text-to-speech (TTS) research has reached a maturity level where excellent speech generation quality is possible on demand, yet two coupled biases persist in reducing perceived quality: accent bias, where models default towards dominant phonetic patterns, and linguistic bias, a misalignment in dialect-specific lexical or cultural information. These biases are interdependent and authentic accent generation requires both accent fidelity and correctly localized text. We present CLARITY (Contextual Linguistic Adaptation and Retrieval for Inclusive TTS sYnthesis), a backbone-agnostic framework to address both biases through dual-signal optimization. Firstly, we apply contextual linguistic adaptation to localize input text to align with the target dialect. Secondly, we propose retrieval-augmented accent prompting (RAAP) to ensure accent-consistent speech prompts. We evaluate CLARITY on twelve varieties of English accent via both subjective and objective analysis. Results clearly indicate that CLARITY improves accent accuracy and fairness, ensuring higher perceptual quality output\footnote{Code and audio samples are available at https://github.com/ICT-SIT/CLARITY.

While large language models (LLMs) offer promising capabilities for automating academic workflows, existing systems for academic peer review remain constrained by text-only inputs, limited contextual grounding, and a lack of actionable feedback. In this work, we present an interactive web-based system for multimodal, community-aware peer review simulation to enable effective manuscript revisions before paper submission. Our framework integrates textual and visual information through multimodal LLMs, enhances review quality via retrieval-augmented generation (RAG) grounded in web-scale OpenReview data, and converts generated reviews into actionable to-do lists using the proposed Action:Objective[\#] format, providing structured and traceable guidance. The system integrates seamlessly into existing academic writing platforms, providing interactive interfaces for real-time feedback and revision tracking. Experimental results highlight the effectiveness of the proposed system in generating more comprehensive and useful reviews aligned with expert standards, surpassing ablated baselines and advancing transparent, human-centered scholarly assistance.

We consider the relative abilities and limitations of computationally efficient algorithms for learning in the presence of noise, under two well-studied and challenging adversarial noise models for learning Boolean functions: malicious noise, in which an adversary can arbitrarily corrupt a random subset of examples given to the learner; and nasty noise, in which an adversary can arbitrarily corrupt an adversarially chosen subset of examples given to the learner. We consider both the distribution-independent and fixed-distribution settings. Our main results highlight a dramatic difference between these two settings: For distribution-independent learning, we prove a strong equivalence between the two noise models: If a class ${\cal C}$ of functions is efficiently learnable in the presence of $η$-rate malicious noise, then it is also efficiently learnable in the presence of $η$-rate nasty noise. In sharp contrast, for the fixed-distribution setting we show an arbitrarily large separation: Under a standard cryptographic assumption, for any arbitrarily large value $r$ there exists a concept class for which there is a ratio of $r$ between the rate $η_{malicious}$ of malicious noise that polynomial-time learning algorithms can tolerate, versus the rate $η_{nasty}$ of nasty noise that such learning algorithms can tolerate. To offset the negative result for the fixed-distribution setting, we define a broad and natural class of algorithms, namely those that ignore contradictory examples (ICE). We show that for these algorithms, malicious noise and nasty noise are equivalent up to a factor of two in the noise rate: Any efficient ICE learner that succeeds with $η$-rate malicious noise can be converted to an efficient learner that succeeds with $η/2$-rate nasty noise. We further show that the above factor of two is necessary, again under a standard cryptographic assumption.

Large Language Models (LLMs) face fundamental challenges in long-context reasoning: many documents exceed their finite context windows, while performance on texts that do fit degrades with sequence length, necessitating their augmentation with external memory frameworks. Current solutions, which have evolved from retrieval using semantic embeddings to more sophisticated structured knowledge graphs representations for improved sense-making and associativity, are tailored for fact-based retrieval and fail to build the space-time-anchored narrative representations required for tracking entities through episodic events. To bridge this gap, we propose the \textbf{Generative Semantic Workspace} (GSW), a neuro-inspired generative memory framework that builds structured, interpretable representations of evolving situations, enabling LLMs to reason over evolving roles, actions, and spatiotemporal contexts. Our framework comprises an \textit{Operator}, which maps incoming observations to intermediate semantic structures, and a \textit{Reconciler}, which integrates these into a persistent workspace that enforces temporal, spatial, and logical coherence. On the Episodic Memory Benchmark (EpBench) \cite{huet_episodic_2025} comprising corpora ranging from 100k to 1M tokens in length, GSW outperforms existing RAG based baselines by up to \textbf{20\%}. Furthermore, GSW is highly efficient, reducing query-time context tokens by \textbf{51\%} compared to the next most token-efficient baseline, reducing inference time costs considerably. More broadly, GSW offers a concrete blueprint for endowing LLMs with human-like episodic memory, paving the way for more capable agents that can reason over long horizons. Code is available at https://github.com/roychowdhuryresearch/gsw-memory.

We study the ability of Transformer models to learn sequences generated by Permuted Congruential Generators (PCGs), a widely used family of pseudo-random number generators (PRNGs). PCGs introduce substantial additional difficulty over linear congruential generators (LCGs) by applying a series of bit-wise shifts, XORs, rotations and truncations to the hidden state. We show that Transformers can nevertheless successfully perform in-context prediction on unseen sequences from diverse PCG variants, in tasks that are beyond published classical attacks. In our experiments we scale moduli up to $2^{22}$ using up to $50$ million model parameters and datasets with up to $5$ billion tokens. Surprisingly, we find even when the output is truncated to a single bit, it can be reliably predicted by the model. When multiple distinct PRNGs are presented together during training, the model can jointly learn them, identifying structures from different permutations. We demonstrate a scaling law with modulus $m$: the number of in-context sequence elements required for near-perfect prediction grows as $\sqrt{m}$. For larger moduli, optimization enters extended stagnation phases; in our experiments, learning moduli $m \geq 2^{20}$ requires incorporating training data from smaller moduli, demonstrating a critical necessity for curriculum learning. Finally, we analyze embedding layers and uncover a novel clustering phenomenon: the top principal components spontaneously group the integer inputs into bitwise rotationally-invariant clusters, revealing how representations can transfer from smaller to larger moduli.

The stereotypical reaching motion of the octopus arm has drawn growing attention for its efficient control of a highly deformable body. Previous studies suggest that its characteristic bend propagation may share underlying principles with the dynamics of a whip. This work investigates whether whip-like passive dynamics in water can reproduce the kinematic features observed in biological reaching and their similarities and differences. Platform-based whipping tests were performed in water and air while systematically varying material stiffness and driving speed. Image-based quantification revealed that the Ecoflex Gel 2 arm driven at 150 rpm (motor speed) reproduced curvature propagation similar to that observed in octopus reaching. However, its bend-point velocity decreased monotonically rather than exhibiting the biological bell-shaped profile, confirming that the octopus reaching movement is not merely a passive whipping behavior. The absence of propagation in air further highlights the critical role of the surrounding medium in forming octopus-like reaching motion. This study provides a new perspective for understand biological reaching movement, and offers a potential platform for future hydrodynamic research.

Recently, Large Language Models (LLMs) have been applied to scientific equation discovery, leveraging their embedded scientific knowledge for hypothesis generation. However, current methods typically confine LLMs to the role of an equation proposer within search algorithms like genetic programming. In this paper, we present SR-Scientist, a framework that elevates the LLM from a simple equation proposer to an autonomous AI scientist that writes code to analyze data, implements the equation as code, submits it for evaluation, and optimizes the equation based on experimental feedback. Specifically, we wrap the code interpreter into a set of tools for data analysis and equation evaluation. The agent is instructed to optimize the equation by utilizing these tools over a long horizon with minimal human-defined pipelines. Empirical results show that SR-Scientist outperforms baseline methods by an absolute margin of 6% to 35% on datasets covering four science disciplines. Additionally, we demonstrate our method's robustness to noise, the generalization of the discovered equations to out-of-domain data, and their symbolic accuracy. Furthermore, we develop an end-to-end reinforcement learning framework to enhance the agent's capabilities.

The creation of systematic literature reviews (SLR) is critical for analyzing the landscape of a research field and guiding future research directions. However, retrieving and filtering the literature corpus for an SLR is highly time-consuming and requires extensive manual effort, as keyword-based searches in digital libraries often return numerous irrelevant publications. In this work, we propose a pipeline leveraging multiple large language models (LLMs), classifying papers based on descriptive prompts and deciding jointly using a consensus scheme. The entire process is human-supervised and interactively controlled via our open-source visual analytics web interface, LLMSurver, which enables real-time inspection and modification of model outputs. We evaluate our approach using ground-truth data from a recent SLR comprising over 8,000 candidate papers, benchmarking both open and commercial state-of-the-art LLMs from mid-2024 and fall 2025. Results demonstrate that our pipeline significantly reduces manual effort while achieving lower error rates than single human annotators. Furthermore, modern open-source models prove sufficient for this task, making the method accessible and cost-effective. Overall, our work demonstrates how responsible human-AI collaboration can accelerate and enhance systematic literature reviews within academic workflows.

Universal sound separation faces a fundamental misalignment: models optimized for low-level signal metrics often produce semantically contaminated outputs, failing to suppress perceptually salient interference from acoustically similar sources. We introduce a preference alignment perspective, analogous to aligning LLMs with human intent. To address this, we introduce MARS-Sep, a reinforcement learning framework that reformulates separation as decision making. Instead of simply regressing ground-truth masks, MARS-Sep learns a factorized Beta mask policy that is steered by a preference reward model and optimized by a stable, clipped trust-region surrogate. The reward, derived from a progressively-aligned audio-text-vision encoder, directly incentivizes semantic consistency with query prompts. Extensive experiments on multiple benchmarks demonstrate consistent gains in Text-, Audio-, and Image-Queried separation, with notable improvements in signal metrics and semantic quality. Our code is available at https://github.com/mars-sep/MARS-Sep. Sound separation samples are available at https://mars-sep.github.io/.

Robotic laparoscopic surgery has gained increasing attention in recent years for its potential to deliver more efficient and precise minimally invasive procedures. However, adoption of surgical robotic platforms remains largely confined to high-resource medical centers, exacerbating healthcare disparities in rural and low-resource regions. To close this gap, a range of solutions has been explored, from remote mentorship to fully remote telesurgery. Yet, the practical deployment of surgical robotic systems to underserved communities remains an unsolved challenge. Humanoid systems offer a promising path toward deployability, as they can directly operate in environments designed for humans without extensive infrastructure modifications -- including operating rooms. In this work, we introduce LapSurgie, the first humanoid-robot-based laparoscopic teleoperation framework. The system leverages an inverse-mapping strategy for manual-wristed laparoscopic instruments that abides to remote center-of-motion constraints, enabling precise hand-to-tool control of off-the-shelf surgical laparoscopic tools without additional setup requirements. A control console equipped with a stereo vision system provides real-time visual feedback. Finally, a comprehensive user study across platforms demonstrates the effectiveness of the proposed framework and provides initial evidence for the feasibility of deploying humanoid robots in laparoscopic procedures.

Optimistic exploration is central to improving sample efficiency in reinforcement learning with human feedback, yet existing exploratory bonus methods to incentivize exploration often fail to realize optimism. We provide a theoretical analysis showing that current formulations, under KL or $α$-divergence regularization, unintentionally bias exploration toward high-probability regions of the reference model, thereby reinforcing conservative behavior instead of promoting discovery of uncertain regions. To address this pitfall, we introduce the General Exploratory Bonus (GEB), a novel theoretical framework that provably satisfies the optimism principle. GEB counteracts divergence-induced bias via reference-dependent reward regulation and unifies prior heuristic bonuses as special cases, while extending naturally across the full $α$-divergence family. Empirically, GEB consistently outperforms baselines on alignment tasks across multiple divergence settings and large language model backbones. These results demonstrate that GEB offers both a principled and practical solution for optimistic exploration in RLHF.

Parallel LLM inference scaling involves sampling a set of $N>1$ responses for a single input prompt. However, these $N$ parallel responses tend to be generated independently from each other, partitioning compute resources and leaving potentially useful information in one generation untapped by others. This is in contrast to response length scaling where past computation is used in all future steps. For higher quality responses and response sets, we propose Bridge to generate interdependent responses in parallel by rethinking batched LLM hidden states as holistic tensors rather than independent slices. With only a small amount (2.8%-5.1%) of new parameters, Bridge improves the relative mean accuracy gains from reinforcement learning with verifiable rewards by up to 39% and boosts consistency of correct responses. Trained once, Bridge scales to any generation width, all with greater performance than independent generations, unlocking a more general mode of parallel scaling that effectively leverages information between sequences, compatible with any post-generation aggregation technique.

Diffusion language models (DLMs) generate tokens in parallel through iterative denoising, which can reduce latency and enable bidirectional conditioning. However, the safety risks posed by jailbreak attacks that exploit this inference mechanism are not well understood. In this paper, we reveal that DLMs have a critical vulnerability stemming from their iterative denoising process and propose a countermeasure. Specifically, our investigation shows that if an affirmative token for a harmful query appears at an intermediate step, subsequent denoising can be steered toward a harmful response even in aligned models. As a result, simply injecting such affirmative tokens can readily bypass the safety guardrails. Furthermore, we demonstrate that the vulnerability allows existing optimization-based jailbreak attacks to succeed on DLMs. Building on this analysis, we propose a novel safety alignment method tailored to DLMs that trains models to generate safe responses from contaminated intermediate states that contain affirmative tokens. Our experiments indicate that the proposed method significantly mitigates the vulnerability with minimal impact on task performance. Furthermore, our method improves robustness against conventional jailbreak attacks. Our work underscores the need for DLM-specific safety research. Our code is available at https://github.com/mdl-lab/dlm-priming-vulnerability.

Heuristic functions are central to the performance of search algorithms such as A-star, where admissibility - the property of never overestimating the true shortest-path cost - guarantees solution optimality. Recent deep learning approaches often disregard admissibility and provide limited guarantees on generalization beyond the training data. This paper addresses both of these limitations. First, we pose heuristic learning as a constrained optimization problem and introduce Cross-Entropy Admissibility (CEA), a loss function that enforces admissibility during training. On the Rubik's Cube domain, this method yields near-admissible heuristics with significantly stronger guidance than compressed pattern database (PDB) heuristics. Theoretically, we study the sample complexity of learning heuristics. By leveraging PDB abstractions and the structural properties of graphs such as the Rubik's Cube, we tighten the bound on the number of training samples needed for A-star to generalize. Replacing a general hypothesis class with a ReLU neural network gives bounds that depend primarily on the network's width and depth, rather than on graph size. Using the same network, we also provide the first generalization guarantees for goal-dependent heuristics.

Counterfactual explanations aim to enhance model transparency by showing how inputs can be minimally altered to change predictions. For multivariate time series, existing methods often generate counterfactuals that are invalid, implausible, or unintuitive. We introduce GenFacts, a generative framework based on a class-discriminative variational autoencoder. It integrates contrastive and classification-consistency objectives, prototype-based initialization, and realism-constrained optimization. We evaluate GenFacts on radar gesture data as an industrial use case and handwritten letter trajectories as an intuitive benchmark. Across both datasets, GenFacts outperforms state-of-the-art baselines in plausibility (+18.7%) and achieves the highest interpretability scores in a human study. These results highlight that plausibility and user-centered interpretability, rather than sparsity alone, are key to actionable counterfactuals in time series data.

Earth observing satellites are powerful tools for collecting scientific information about our planet, however they have limitations: they cannot easily deviate from their orbital trajectories, their sensors have a limited field of view, and pointing and operating these sensors can take a large amount of the spacecraft's resources. It is important for these satellites to optimize the data they collect and include only the most important or informative measurements. Dynamic targeting is an emerging concept in which satellite resources and data from a lookahead instrument are used to intelligently reconfigure and point a primary instrument. Simulation studies have shown that dynamic targeting increases the amount of scientific information gathered versus conventional sampling strategies. In this work, we present two different learning-based approaches to dynamic targeting, using reinforcement and imitation learning, respectively. These learning methods build on a dynamic programming solution to plan a sequence of sampling locations. We evaluate our approaches against existing heuristic methods for dynamic targeting, showing the benefits of using learning for this application. Imitation learning performs on average 10.0\% better than the best heuristic method, while reinforcement learning performs on average 13.7\% better. We also show that both learning methods can be trained effectively with small amounts of data.