Generating chemically valid 3D molecular conformations is critical for computational drug discovery. Classical diffusion-based models like GeoLDM perform well but require hundreds of steps, making large-scale in silico screening impractical. Recent efforts on few-step molecular generation have accelerated this process to 12-50 steps, but they often largely sacrifice sample stability. In this work, we present FlashMol, an ultra-fast molecule generative model producing high-quality molecular conformations in as few as 4 steps. To achieve this, we adapt distribution matching distillation (DMD) -- a reverse KL-divergence minimization objective -- to the molecular domain for effective distillation. Considering the local minimization behavior of DMD, we respace the molecule generation timesteps, providing the generator with much better initialization and enables effective distillation. Additionally, to mitigate the mode-seeking behavior of DMD and improve diversity, we further regularize it with a Jensen-Shannon divergence term, which incorporates the mean-seeking behavior of the forward KL divergence. Extensive experiments on QM9 and GEOM-DRUG datasets demonstrate that FlashMol matches and even surpasses the original 1000-step teacher: by applying respacing to the GeoLDM teacher (GeoLDM-respaced), we achieve comparable results to the 1000-NFE teacher with as few as 16 NFEs. After further distillation, FlashMol is able to sample with only 4 NFEs while maintaining high molecular quality, achieving up to 250$\times$ acceleration compared to the GeoLDM teacher.
Skill-augmented agents increasingly rely on large reusable skill libraries, but retrieving relevant skills is not the same as presenting usable context. Existing methods typically return atomic skills or dependency-aware bundles whose internal roles remain implicit, leaving the agent to infer the execution entry point, support skills, visible requirements, and failure-avoidance guidance. We introduce Group of Skills (GoSkills), an inference-time group-structured retrieval method that changes the agent-facing retrieval object from a flat skill list to a compact, role-labeled execution context. GoSkills builds anchor-centered skill groups from a typed skill graph, expands support groups through a group graph, bottlenecks the selected group plan into a bounded set of atomic skill payloads, and renders a fixed execution contract with Start, Support, Check, and Avoid fields, without changing the downstream agent, skill payloads, or execution environment. Across SkillsBench, ALFWorld, and Terminal-Bench 2 with six backbones, GoSkills attains the highest mean reward or accuracy among the evaluated methods in every benchmark-backbone setting, and matched controls show higher reward and fewer agent turns with group selection than with fixed-contract or post-hoc role-annotation alternatives.
Predicting whether a protein sequence possesses a biochemical property of interest is important for understanding and designing protein functionality. Modern protein sequence databases provide valuable training data for functional prediction, but they exhibit survivorship bias: natural selection favors variants that support organismal survival while eliminating many others before they can be observed, complicating classifier training. This setting motivates positive-unlabeled (PU) learning, which treats observed functional sequences as positive and unobserved sequences as unlabeled rather than nonfunctional. However, existing protein PU methods do not account for cross-sequence heterogeneity in evolutionary accessibility: sequences one nucleotide mutation away from a common variant arise far more readily than distant ones. We introduce PULSE: PU Learning with Surveillance and Evolution, a framework that trains a probabilistic classifier for a target protein function using an evolution-aware observation likelihood. When considering functionality of proteins with high genomic sequence coverage, PULSE combines a probabilistic classifier that estimates whether an amino acid sequence performs the target function with an evolutionary model that estimates the likelihood that its encoding nucleotide sequences emerge and are detected based on mutational accessibility, parental prevalence, and surveillance e!ciency. By coupling these two components, PULSE treats the absence of an evolutionarily accessible sequence as stronger evidence of non-functionality than the absence of one unlikely to arise. Tested on eleven tasks focusing on multiple Spike protein properties spanning five SARS-CoV-2 variants, PULSE ranks first or second on all of them, and achieves the best overall performance among all methods, including standard PU learning, one-class classification methods, and protein language models.
Large Language Models (LLMs) demonstrate strong capability in solving scientific and mathematical problems, yet they struggle to produce valid and challenging novel problems -- an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-backed hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin. Our code is available at https://github.com/Miaow-Lab/VHG.
Despite an ever-increasing interest in topological deep learning models that target higher-order datasets, there is no consensus on how to evaluate such models. This is exacerbated by the fact that topological objects permit operations, such as structural refinements, that are not appropriate for graph data. In this work, we extend MANTRA, a benchmark dataset containing manifold triangulations, to a larger class of manifolds with more diverse homeomorphism types. We show that, contrary to prior claims, graph neural networks can saturate the benchmark, while higher-order message passing methods lag behind. However, we find that this is contingent on the right representation and feature assignment, emphasizing their importance in baseline models. Hence, we present a novel evaluation protocol based on representational diversity and triangulation refinement. Even when including refinement-like changes in the training data, existing models fail to generalize beyond the combinatorial structure of the data. This points towards a research gap in developing models that understand topological structure independent of scale. Our work thus provides the necessary scaffolding to evaluate future models and enable the development of topology-aware inductive biases.
We study Nash equilibrium learning in partially observable Markov games (POMGs), a multi-agent reinforcement learning framework in which agents cannot fully observe the underlying state. Prior work in this setting relies on centralization or information sharing, and suffers from sample and computational complexity that scales exponentially in the number of players. We focus on a subclass of POMGs with independent state transitions, where agents remain coupled through their rewards, and assume that the underlying fully observed Markov game is a Markov potential game. For this class, we present an independent learning algorithm in which players, observing only their own actions and observations and without communication, jointly converge to an approximate Nash equilibrium. Due to partial observability, optimal policies may in general depend on the full action-observation history. Under a filter stability assumption, we show that policies based on finite history windows provide sufficient approximation guarantees. This enables us to approximate the POMG by a surrogate Markov game that is near-potential, leading to quasi-polynomial sample and computational complexity for independent Nash equilibrium learning in the underlying POMG.
In this paper, we introduce INEUS, a meshfree iterative neural solver for partial integro-differential equations (PIDEs). The method replaces the explicit evaluation of nonlocal jump integrals with single-jump sampling and reformulates PIDE solving as a sequence of recursive regression problems. Like Physics-Informed Neural Networks (PINNs), INEUS learns global solutions over the entire space-time domain, yet it offers a more efficient treatment of nonlocal terms and avoids the computationally expensive differentiation of full PIDE residuals. These features make INEUS particularly well suited for high-dimensional PDEs and PIDEs. Supported by a contraction-based convergence proof for linear PIDEs, our numerical experiments show that INEUS delivers accurate and scalable solutions for various high-dimensional linear and nonlinear examples.
We propose Stochastic Interpolant Prior for Speech (SIPS), a plug-and-play framework for speech enhancement and separation that augments pretrained predictors with a generative clean speech prior. Building on stochastic interpolants, we decompose the sampling dynamics into a predictor-induced drift and a score-based denoising component. The predictor guides sampling toward a task-consistent estimate, while a score model trained exclusively on clean speech promotes perceptual naturalness. Unlike prior hybrid approaches, which typically rely on architecture-specific conditioning and are tied to particular predictors or degradation settings, SIPS employs a degradation-agnostic prior that can be reused across predictors and additive degradation tasks without retraining. We evaluate SIPS on speech enhancement and separation using recent predictors such as SEMamba and FlexIO. Across both tasks, SIPS consistently improves non-intrusive speech quality metrics while maintaining competitive performance on reference-based distortion measures and word error rate. For speech enhancement, these improvements are further supported by a subjective listening test.
Addressing missing modalities is an important challenge in multimodal image analysis and often relies on complex architectures that do not transfer easily to different datasets without architectural modifications or hyperparameter tuning. While most existing methods tackle this problem in feature space by engineering representations that are robust to missing inputs, we instead operate in weight space. We propose LARGO, a hypernetwork that compresses the $2^N-1$ dedicated missing-modality models into a single network by modelling the convolutional weights using the Canonical Polyadic (CP) tensor decomposition. Extensive experimental validation on BraTS 2018 (4 modalities, 15 scenarios) and ISLES 2022 (3 modalities, 7 scenarios) shows that our method ranks first in 47 out of 52 configurations, achieving average Dice improvements of +0.68$\%$ and +2.53$\%$ over state-of-the-art baselines (mmFormer, M$^{3}$AE, ShaSpec, SimMLM). A proof-of-concept experiment on avMNIST suggests that LARGO may extend beyond medical imaging to heterogeneous non-medical modalities. Code is available at https://github.com/nielsvyncke/LARGO.
In multiclass softmax cross-entropy, the full generalized Gauss-Newton (GGN) curvature couples all output logits through the softmax covariance, making curvature-vector products harder to scale as the number of classes grows. We show that the standard multiclass GGN can be decomposed exactly into a true-vs-rest term and a positive semidefinite within-competitor covariance term. Fast Gauss-Newton (FGN) retains the first term and drops the second, yielding a positive semidefinite under-approximation of the multiclass GGN that is exact for binary classification. The derivation uses an exact true-vs-rest scalar-margin representation of softmax cross-entropy: the loss and gradient are unchanged, and the approximation enters only at the curvature level. Exploiting the FGN curvature structure, the damped update can be written as an equivalent whitened row-space system with one row per mini-batch example. We solve this system matrix-free by conjugate gradient using Jacobian-vector and vector-Jacobian products of the scalar margin map. Targeted mechanism experiments show how the dropped covariance grows with competitor dispersion and how damping can reduce its effect on the resulting steps. In frozen-feature head optimization and end-to-end CIFAR-100 training, FGN achieves similar test accuracy to a full GGN baseline with faster updates.
Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.
Large language models (LLMs) are increasingly applied to complex problems, from automated code debugging to science itself. Such problems require generating plausible hypotheses from sparse, noisy, and ambiguous data and updating them as more information arrives. Yet how closely LLMs approach optimal inductive reasoning remains unclear. We study this question in the classic number game, in which a learner infers a concept from a few ambiguous positive integers: $\{2, 16, 4, 8\}$, for example, is explained both by ``powers of 2'' and by ``numbers up to 20.'' We probe the concepts LLMs hypothesize from the observed examples and use their behavioral patterns to understand how they generate and update hypotheses. We measure the LLM posterior over hypotheses in three tasks (prediction, evaluation, and generation) and compare LLM behavior with a Bayesian model and with humans. LLMs favor narrower hypotheses than humans and show a robust evaluation--generation gap: in evaluation they select hypotheses more consistent with the observed examples, whereas in generation they propose narrower hypotheses that are more often rules. They also update poorly: as examples accumulate, their accuracy declines while their confidence grows. Thinking reduces, and can even reverse, the gap; it also broadens the hypotheses LLMs favor, raising accuracy while lowering confidence. Whereas prior studies have shown that LLMs are good hypothesis proposers, our results and further comparisons with the Bayesian model show that LLM inductive reasoning lacks coherence across the three tasks, suggesting that LLMs do not maintain a stable representation shared across tasks that make the same demand. Our results highlight a limitation of LLMs in inductive inference, particularly in settings such as science, where hypotheses must go beyond the data.
We investigate the ability of transformers to perform in-context reinforcement learning (ICRL), where a model must infer and execute learning algorithms from trajectory data without parameter updates. We show that a linear self-attention transformer block can provably implement policy-improvement methods, including semi-gradient SARSA and actor-critic, via explicit parameter constructions. Beyond existence, we design a teacher-mimicking training procedure, analyze its gradient-flow dynamics, and establish the first convergence guarantee in the ICRL literature: under suitable richness conditions on the training MDP distribution, gradient flow converges locally and exponentially to an optimal parameter manifold corresponding to the desired RL update. Empirically, training transformers on randomly generated tabular MDPs confirms these predictions: the learned models recover the parameter structure of our explicit constructions and, when deployed on unseen MDPs, deliver strong in-context control performance. Together, these results illuminate how transformer architectures internalize and execute classical reinforcement learning algorithms in context, bridging mechanistic understanding and training dynamics in ICRL.
We study fixed-confidence best arm identification in generalized linear bandits under a hybrid feedback model: at each round, the learner may query either (i) absolute reward feedback from a single arm or (ii) relative (dueling) feedback from an arm pair, both governed by generalized linear models. We introduce a likelihood-ratio--based confidence sequence that unifies heterogeneous generalized linear observations and yields an explicit ellipsoidal confidence set under a self-concordance assumption. Building on this confidence set, we propose a hybrid Track-and-Stop algorithm that adaptively allocates queries by tracking a minimax-optimal design over a joint action space of arms and pairs. We establish $\delta$-correctness and provide almost-sure asymptotic upper bounds on the stopping time. We further extend the framework to a cost-aware setting that accounts for heterogeneous acquisition costs across feedback modalities. Empirical experiments demonstrate that the proposed algorithms significantly improve sample efficiency over baseline methods.
Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We study these attacks through the lens of a reconstruction-concealment tradeoff: the transformed input must sufficiently obscure harmful intent while preserving enough information for the victim model to recover the underlying request. Through a reconstruction and concealment analysis of representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants, constructed by randomly removing characters from the original query, achieve a better balance, motivating their use as the basis for our attack construction. Building on this observation, we propose concealment-aware variant construction, which selects character-removed variants with low harmful keyword/phrase alignment and high mutual diversity, and instantiates them through five prompting strategies that distribute the transformed information across text and image channels. We further introduce keyword/phrase-related distractor images that depict the harmful concept in diverse contexts, providing more effective distraction than generic distractor images. Experiments across diverse closed-source and open-source MLLMs show that the proposed strategies substantially outperform the evaluated baselines. Code and reproduction instructions are available at https://github.com/Farhamdur/CRJ.
Generative compressed sensing uses the range of a pretrained generator as a nonlinear model for recovering structured signals from limited measurements. We study a conditional version of this problem for image recovery from subsampled Fourier measurements using prompt-conditioned generative models. Our framework separates two roles of conditioning: the prompt used to design the sampling distribution and the prompt used to define the recovery model. For ReLU and Lipschitz conditional generators, we prove stable recovery bounds showing that prompt-matched Christoffel sampling retains the same Christoffel complexity constant as existing near-optimal generative compressed sensing theory, while prompt mismatch incurs an explicit compatibility penalty. Experiments with Stable Diffusion show that prompts meaningfully reshape Christoffel sampling distributions and influence image recovery. Overall, our results suggest treating prompts as design variables for sensing, approximation, and recovery, opening new directions for the practical use of conditional generative models in compressed sensing, where conditioning, sampling, and optimization must be considered jointly.
Large language model (LLM) agents are increasingly used to operate browsers, files, code and tools, making personal assistants a natural deployment target. Yet personal agents face a privacy-cost-capability tension: cloud models execute multi-step workflows well but expose sensitive intermediate context to external APIs, while local models preserve privacy but remain less reliable. Both settings also pay repeatedly for long skill prompts and growing histories. We propose constant-context skill learning, a context-to-weights framework for recurring agent workflows: reusable procedures are learned in lightweight task-family modules, while inference conditions only on the current observation and a compact state block. A deterministic tracker renders this state block from task progress and supplies aligned subgoal rewards, so each module can be trained with step-level SFT and refined through online RL. Across ALFWorld, WebShop, and SciWorld, our agents achieve strong performance across Qwen3-4B, Qwen3-8B, Qwen3-14B, and Llama-3.1-8B. With Qwen3-8B, SFT+RL reaches 89.6% unseen success on ALFWorld, 76.8% success on WebShop, and 66.4% unseen success on SciWorld. They match or exceed strong published agent-training results while reducing prompt tokens per turn by 2-7x relative to controlled ReAct prompting baselines, showing that procedural context can be moved from prompts into weights. Code is available at https://github.com/haoyangxie/constant-context-skill-learning.
This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture overt forms of the phenomenon but leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sycophancy should be defined not by agreement or factual incorrectness alone, but by cue-driven degradation of epistemic integrity. To clarify this boundary, we propose a three-condition decision rule requiring (1) a user cue, (2) a model response that shifts toward that cue, and (3) a resulting degradation of epistemic integrity. Building on this definition, we also introduce a taxonomy consisting of alignment targets, mechanisms, and severity to characterize cases identified as sycophantic. An initial 48-item annotation pilot provides preliminary evidence that the framework can be operationalized using structured guidelines, although broader empirical validation remains necessary. The paper concludes by discussing implications for alignment evaluation, including boundary-aware assessment, structured rubrics, and mitigation strategies, while situating these proposals alongside alternative views of sycophancy.
Reinforcement learning from human feedback (RLHF) typically assumes a static or non-strategic reward model (RM). In iterative deployment, however, the policy generates the data on which the RM is retrained, creating a feedback loop. Building on the Stackelberg game formulation of this interaction, we derive an analytical decomposition of the policy's true optimization gradient into a standard policy gradient and a parameter-steering term that captures the policy's influence on the RM's future parameters. We show that standard iterative RLHF, which drops this steering term entirely, suffers from alignment collapse: the policy systematically exploits the RM's blind spots, producing low-quality, high-reward outputs whose feedback reinforces the very errors it exploits. To mitigate this, we propose foresighted policy optimization (FPO), a mechanism-design intervention that restores the missing steering term by regularizing the policy's parameter-steering effect on RM updates. We instantiate FPO via a scalable first-order approximation and demonstrate that it prevents alignment collapse on both controlled environments and an LLM alignment pipeline using Llama-3.2-1B.
Transformers are effective at inferring the latent task from context via two inference modes: recognizing a task seen during training, and adapting to a novel one. Recent interpretability studies have identified from middle-layer representations task-specific directions, or task vectors, that steer model behavior. However, a lack of rigorous foundations hinders connecting internal representations to external model behavior: existing work fails to explain how task-vector geometry is shaped by the training distribution, and what geometry enables out-of-distribution (OOD) generalization. In this paper, we study these questions in a controlled synthetic setting by training small transformers from scratch on latent-task sequence distributions, which allows a principled mathematical characterization. We show that two inference modes can coexist within a single model. In-distribution behavior is governed by Bayesian task retrieval, implemented internally through convex combinations of learned task vectors. OOD behavior, by contrast, arises through extrapolative task learning, whose representations occupy a subspace nearly orthogonal to the task-vector subspace. Taken together, our results suggest that task-vector geometry, training distributions, and generalization behaviors are closely related.