Computer Science & Mathematics

Preference optimization can learn spurious correlations, and augmenting the preference data with ties (equally preferred responses) can reduce spurious reliance. However, ties differ in their corrective effect, making their selection for augmentation important. For log-linear policies, we derive each tie's influence on spurious reliance and show that taking the largest predicted reductions is first-order optimal among subsets of equal size. Computing these influences requires knowledge of the spurious features and the curvature of the training objective, which is often unavailable or expensive to estimate. To this end, we explore using the magnitude of a tie's implicit reward margin and characterize the conditions under which it is a useful proxy for predicted influence. To test the theory and its qualitative transfer to neural models, we estimate influence with a counterfactual proxy for spurious reliance and use a generalized Gauss--Newton approximation to curvature for language models. Across log-linear policies, neural models, and large language models, our selection rules consistently achieve comparable reductions in spurious reliance with fewer ties than uniform selection.

Phase transitions can generate intricate spatial organization invisible to spatially averaged probes, yet quantitative measures of this organization remain limited. Here we develop "entropy imaging," an information-theoretic technique that harnesses lossless compression to quantify spatial complexity, predictability, and order in quantum materials. Using this approach, emergent spatial textures can be imaged and converted into an experimentally-accessible observable: the computable information density (CID), a compression-based measure of information complexity that approaches the Shannon entropy rate in the large-system limit. Applied to millikelvin microwave impedance microscopy of graphene, the method reveals a pronounced maximum in the CID near quantum Hall transitions, a quantitative measure of the information-rich spatial organization of the critical landscape. A disordered Hofstadter model reproduces this behavior and shows that the CID can identify the quantum critical point from a sequence of images, without ensemble averaging over disorder realizations. The same framework tracks fractional quantum Hall transitions and resolves isolated fractional puddles that coexist with a disordered background. By contrast, the trivial band-edge crossover in a MoSe2/WS2 moiré heterobilayer exhibits qualitatively different behavior, suggesting that the pronounced CID enhancement observed at the quantum Hall critical point is not simply a generic feature of a conductive-insulating crossover. We further show that the geometric information encoded in CID is not reducible to a correlation length, and it can even distinguish spatial textures with identical multifractal statistics. Together, these results establish lossless compression as a framework for quantifying emergent spatial organization in quantum materials, with portability across imaging modalities and physical systems.

Deep research, or any use of large language models (LLMs) to read a large document collection and produce insights, requires synthesizing information from different sources. In this work, we find that LLMs produce long-form research reports that read as confident syntheses but which fail to accurately reflect the underlying documents. We call this failure smoothing. In collapsing diverse and potentially conflicting sources into a single narrative, the model discards the evidential "jaggedness" of its corpus and asserts conclusions more decisive than the citations can support. We introduce a claim level scoring pipeline, SmoothScore, to quantify this effect. Our pipeline extracts key conclusive claims from model generated reports, then finds evidence relevant to each claim and clusters that evidence along high level aspects. Finally, each claim is scored with respect to the evidence to quantify the degree of smoothing done by the agent. Across 7 systems and 4 datasets, we find that agents consistently produce stronger claims than warranted.

If $G$ and $H$ are two cubic graphs, then an $H$-coloring of $G$ is a mapping $f: E(G)\rightarrow E(H)$, such that for every $v\in V(G)$ there is a vertex $x\in V(H)$, such that $f(\partial_G(v))=\partial_H(x)$. If $G$ admits an $H$-coloring then it is common to write $H\prec G$. In this paper, for a fixed connected cubic graph $H$, we consider a decision problem, which we call $H$-coloring. In this problem, on the input one is given a connected cubic graph $G$. The goal is to decide whether $H\prec G$. A conjecture by Hakobyan and the author states that $H$-coloring is polynomial time solvable, if $H$ does not admit a $P_{12}$-coloring. Here $P_{12}$ is the unique cubic graph obtained from the Petersen graph by expanding one of its vertices into a triangle. In this paper, we disprove this conjecture under $P\neq NP$, by showing that $S_6$-coloring is NP-complete. Here $S_6$ is the unique cubic graph on six vertices having exactly one bridge. Our second result is an algorithm that solves $H$-coloring problem in time $O^*(c^n)$, where $H$ is a simple connected cubic graph, $n$ is the number of vertices of the input cubic graph $G$ and $c=\sqrt{8}\approx 2.8284$.

Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions. Learned critics provide value estimates for intermediate states, but these estimates can be inaccurate, while estimating intermediate values with repetitive continuation samples from points within a trajectory incurs substantial costs. To mitigate these issues, we introduce FOCUS, an RL method that uses selective branching to help the critic network learn its value estimator. The rollout controller facilitates active learning for the critic by sampling continuations from states that are expected to most reduce the critic's estimation error. These refined value estimates then score the segment between consecutive states with a TD($λ$) advantage, while the sampled evidence supplies training targets for the critic. FOCUS supports long-horizon agentic reinforcement learning through a modular context-management interface that decouples context management from rollout allocation. Our experiments show that using compaction-induced segment boundaries for value estimation and credit assignment, FOCUS outperforms GRPO and VinePPO with matched per-step generation budgets on legal reasoning, deep search, and code localization benchmarks, improving held-out accuracy at Qwen3.5-9B scale by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs. 48.2) and by 4.1 points on DeonticBench (56.5 vs. 52.4).

Many academic disciplines (in particular, mathematics and the theoretical sciences) face difficult questions about how to adapt their practices to the rise of highly capable artificial intelligence (AI). Before AI, a well-written research paper with a strong theoretical result acted as a reliable signal that its authors understood their field in depth. Now that AI is widely accessible, such signals have become cheaper to produce. As a result, they no longer reliably sort academics by their level of understanding. This has direct consequences for academic hiring. The age of AI raises a question for each side of the academic job market. Should an early-career researcher invest in a deep understanding of their field, or instead use AI to build a longer resume with many papers? Should a recruiter favor applicants with fewer papers who pass an interview aimed at probing understanding, or insist on both publication volume and a pass? In this paper, we develop a game-theoretic model of this two-sided dilemma. We classify all pure-strategy equilibria that can arise in this game and use learning simulations to identify the behaviors that populations of applicants and recruiters may learn to adopt. Our analysis provides a framework for reasoning about how academic hiring can adapt to AI and for evaluating practices that preserve the value of deep understanding.

The transition metal trichalcogenides (TMTs) are an interesting class of layered 2D materials that have strong anisotropic electrical and optical properties. We report on the polarization-dependent nonlinear optical response of exfoliated crystals at the third harmonic using non-resonant excitation with a 1560 nm femtosecond laser. All members of the TMT crystal family explored (ZrS3, HfS3, TiS3, and NbS3) display large anisotropy in their polarized third harmonic generation signal, with anisotropy ratios as large as chi_11/chi_22 ~ 6.5. From the expected nonlinear dependence of these monoclinic and triclinic crystals we extract the relative nonlinear susceptibility coefficients. We find that the anisotropy ratio is largest for ZrS3 and decreases across the family, a trend we associate with the proximity of the third harmonic energy to the strongly b-polarized absorption edge of each material. NbS3, the only triclinic crystal, displays a pronounced anisotropy in thick flakes that is strongly reduced in the thinnest flakes, which we attribute to a partial relaxation of the Peierls-type Nb-Nb dimerization that underlies the triclinic distortion. The large anisotropic nonlinear response in this crystal family sets it apart from other 2D materials and makes it particularly attractive for further investigation.

Diffusion large language models (dLLMs) can generate multiple tokens in parallel, but often require many decoding rounds to complete a response. Efficient generation depends not only on which tokens are easy to predict, but also on how committing to them helps predict the remaining positions. A highly predictable token may provide little useful context, while revealing a less predictable token can help several other positions become predictable sooner. We propose dSpark, a new schedule distillation-based framework for learning when to commit tokens for faster diffusion language generation. To achieve this, we construct unmasking schedules that prioritize tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier in decoding. We then distill these schedules into the model to enable faster parallel sampling, with a KL-divergence regularizer to help preserve generation quality. We extensively evaluate our approach across mathematical reasoning and code generation tasks. On GSM8K, our method achieves remarkable 10.12$\times$ and 5.12$\times$ speedups over the various baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining superior task performance. Our code is available at https://github.com/VILA-Lab/dSpark.

Feed-forward 3D reconstruction models such as VGGT regress cameras and geometry in a single forward pass, but they report no uncertainty for the camera poses they predict: an incorrect relative pose looks no different from a correct one. We ask whether the frozen model nonetheless knows when its pose is wrong. On HardMatch, a benchmark of 1,000 deliberately difficult image pairs on which VGGT fails on 35% of pairs, a linear probe on VGGT's frozen features predicts pose failure on held-out categories with ROC-AUC 93.5%, compared with 79.7% for VGGT's own point-map confidence and 77.7% for a classifier trained on all of VGGT's outputs. The probe catches failures VGGT is confident about, lets a downstream system abstain on the pairs it distrusts, and transfers without retraining to rotation failures on PhotoTourism and ScanNet-1500, beyond what co-visibility explains. This knowledge is absent from the per-image encoder: it is computed partway through the layers that fuse the two views, and the same holds for a second model, $π^3$. In essence, we find that VGGT knows more about its own failures than it reports.

We present a robust Speech and Audio Quality Metric (SAQM) that merges features of traditional quality metrics with a CNN trained on a large amount of data using contrastive learning methods. To overcome the problem of limited amounts of human-labeled training data, we propose two pretraining strategies relying on pseudo-labels. One uses an established objective metric PEAQ-2f (based on ITU-R BS.1387) as surrogate loss, the other is based on contrasting a wide range of coding modes by a rule-of-thumb ordering based on expected perceived quality. As a key novelty, our approach lets us generate saliency maps by propagating gradients to identify which feature has the greatest impact on the final metric outcome and enables our system to provide insights regarding the type and extent of distortions present in the tested audio signal. The resulting (SAQM) operates at low computational complexity, performs comparably to (SOTA) metrics on speech while still performing well on general audio content.

Dynamic metasurface antennas (DMAs) have emerged as a promising technology for realizing large-scale antenna arrays in next-generation wireless networks owing to their ability to enable dense and energy-efficient antenna architectures. However, jointly optimizing digital beamforming and DMA element weights under time-varying channel conditions remains challenging due to their coupled effects on the received signal and interference powers. We formulate this problem to maximize time-averaged energy efficiency (EE) under time-varying channel conditions, subject to transmit power constraints, Lorentzian phase limitations, and minimum rate requirements. To enable long-term planning and improve sample efficiency compared with model-free reinforcement learning, we develop a world model framework that learns latent system dynamics from interactions with the environment and jointly optimizes digital beamforming and DMA element weights through imagined trajectories. Furthermore, we incorporate feasibility-preserving action mappings and safety cost signals to ensure constraint satisfaction during policy learning. Simulation results demonstrate that the proposed framework improves the time-averaged EE by approximately 20% compared with the deep Q-network baseline, while achieving higher sample efficiency and more reliable constraint satisfaction.

The security region (SR) specifies a joint feasible region in the operational space involving various installations within the power system. It enables dynamic allocation of capacity among these energy units, thereby unlocking additional grid potential for power delivery and mitigating congestion. By constructing a convex hull from a set of vertices to form a polytope, the SR can be effectively represented with an acceptable accuracy loss. However, the SR computation inevitably generates many low-quality vertices, which contribute little to the overall hypervolume and can be discarded. By introducing SR compression, the SR data size can be reduced, which in turn facilitates the memory-efficient computation of SR and time-efficient solution of SR-related optimization problems. Accordingly, this paper proposes a vertex pruning approach that mixes local and backward search policies to achieve effective SR compression. In particular, the backward search strategy allows previously pruned vertices to be recycled when they are re-identified as critical. Subsequently, we incorporate this approach to enhance computational efficiency in the aforementioned applications. Numerical validations are conducted on two realistic medium-voltage Dutch grid cases. The results demonstrate at most 1.65% hypervolume loss with over 96.0\% SR data size reduction after applying the proposed compression approach. Empowered by this compression, the 8-dimensional (8D) SR becomes computationally tractable. Moreover, in relevant grid-aware computation challenges, the solving time can be reduced by at least 52.9% with at most an 1.36% optimality loss. All above validate both the efficacy and values of the proposed SR compression approach.

Editing multi-camera concerts requires coordinated decisions about musical relevance, visual quality and cut timing. We present a modular framework that combines instrument activity from multitrack audio, visible instruments and image sharpness with beat-based timing and probabilistic camera selection. An interpretable scoring model links these components to the choice of views over complete shot intervals. Component tests and an exploratory expert study on a four-camera concert yield mixed assessments across the editing variants. The findings motivate extending local audiovisual relevance with contextual and sequence-level criteria. The architecture provides a basis for testing these extensions and studying how component behaviour affects the quality of complete edits.

A robot policy needs to understand what a task is asking for in a scene, and how to move the robot to accomplish that task. To achieve these two skills, we introduce Zero-real Agent Policy (ZAP), a recipe that teaches a small off-the-shelf vision-language model to control a robot via a simple tool harness. We use a pretrained 27B parameter model to bootstrap world understanding and reasoning, followed by supervised fine-tuning and GRPO on simple simulated tasks to teach it how to move the robot. Without any real-world robot data during training, ZAP reaches 96.7% success on real-world tasks, matching GPT-6 Astra in our harness and outperforming the MolmoAct2 VLA (33.3%), while running 3x faster and 17x cheaper than GPT-6 Astra. We demonstrate that ZAP also generalizes to completely new objects, goals, strategies, and language by seating a test tube in a rack (and choosing a different hole in the rack when prompted), handing a test tube between arms mid-air, drawing an X on a tic-tac-toe board, and playing chess. These goals and types of manipulation don't appear in our training data. Furthermore, ZAP also transfers to an unseen embodiment. On a single-arm Franka, ZAP reaches 55.9% on RoboLab-120, 4.9 percentage points ahead of the leading VLA/WAM baseline.

Browser agents do not just need more practice. They need to practice the right things. We introduce MiniWeb, a realistic miniature of the web comprising 65 interconnected websites across 15 domains, where agents navigate, search, and transact as people do. Its 369 human-annotated tasks cover everyday workflows and label 37 reusable interaction primitives grouped into 7 skill families, allowing us to diagnose which skills fail rather than only whether a task succeeds. Benchmarking ten open-weight agents, we find that drag & gesture is consistently the weakest family, with no agent solving more than 24% of its instances. To address this gap, we introduce MiniWebGen, a framework that autonomously generates verified, skill-centric training data from MiniWeb. Fine-tuning Qwen3.5-4B on just 3.1K verified trajectories improves single-primitive success by 24.9pp across 13 unseen websites, with gains across every skill family. However, these trajectories teach short-horizon interactions, which do not directly translate to full-task success. We build on this result with MiniWebAgent, which pairs the trained executor with a planner that decomposes full tasks into the skill-level steps it learned to perform. This approach improves task success by 19.4pp points on unseen MiniWeb sites and transfers to external benchmarks, with gains of 9.3pp on WebArena-Lite, 16.1pp on Online-Mind2Web, and 20.0pp on WebVoyager, without training on their trajectories. We provide a simple recipe for compact browser agents: practice reusable skills directly, then compose them to solve unfamiliar tasks.