Papers

In so-called "Industry 4.0," integrated sensors, analytics tools, and other digital technologies supplement industrial workflows, potentiating enhanced visibility and automation. Prognostics and health management (PHM) and natural language processing (NLP) practitioners alike have explored the application of large language models (LLMs) to industrial asset management, but many prior works' focus on exam-like benchmarks has limited their assessment of LLMs' ability to automate practical operations. Herein, we introduce CoffeeBench, a fault diagnostics task for LLM agents, using a simulated appliance. The LLMs we evaluate struggle to consistently identify faults from the sensor information alone, but the best-performing ones can iteratively refine the fault-option space to eventually obtain a successful diagnosis. Our subsequent analyses, therefore, focus on: differentiating models by diagnostic efficiency (RQ1); examining the lagging performance of the three worst models (RQ2); and identifying viable heuristics for selective prediction (RQ3), including that initial diagnostic uncertainty and LLM-judged soundness of reasoning are high-signal predictors of diagnostic success. In addition to our benchmark results (Section 3.1), we provide practical insights for improving the performance of the open-weight models on CoffeeBench, such as by explicitly tracking episode state on their behalf to reduce timeouts (Section 3.2) or applying a deliberation loop to refine diagnostic hypotheses (Section 3.3).

We present Kandinsky S2V 1.0, an open family of speech-to-video models for controllable audio-driven human animation, comprising Kandinsky S2V 1.0 Lite (2B parameters) and Kandinsky S2V 1.0 Pro (21B parameters). From a single reference image, a speech waveform, and a text prompt, the models generate long-form 25 FPS video in which lip motion, expression, pose, and gesture follow the speech. The architecture draws text and audio conditioning from one Qwen2.5-Omni model and adds a reference-aware three-scale visual memory. Training starts from the Kandinsky 5.0 Video foundation models [1] and combines mixed reference-memory layouts with an area-adaptive face loss. The training corpus contains 1.17M single-speaker clips selected by active-speaker detection, with captions generated directly in English and Russian. A side-by-side human evaluation against the strongest open-source baseline and two proprietary commercial models shows that Kandinsky S2V 1.0 Pro performs comparably to or better than these models on average and is also preferred for lip synchronization. On common speech to video benchmarks, Kandinsky S2V 1.0 Lite and Kandinsky S2V 1.0 Pro rank among the top performing models according to mean rank over five reported metrics. Beyond the base models, we provide {\pi}-Flow distilled variants of both Kandinsky S2V 1.0 Lite and Kandinsky S2V 1.0 Pro. The distilled Kandinsky S2V 1.0 Pro model generates video in 16 sampling steps without classifier-free guidance, achieving a 5.4 times end-to-end speedup over the base model, while matching or exceeding it on nearly all reported benchmark metrics. To accelerate open research and deployment in speech-to-video generation, we release the code, model checkpoints, and diffusers integration under the MIT license.

The electric power grid is a critical infrastructure system subject to different types of failures which can have devastating impacts on society. One framework to study the resilience of the grid is the N-k interdiction model, which introduces an attacker and defender model to identify grid vulnerabilities. However, traditional N-k interdiction models characterize the power grid with steady-state assumptions. One key aspect missing from this analysis is the behavior of the grid as it evolves from the pre-failure grid state and post-failure grid state. During such a transition period, automated protection systems, transient instabilities, and other phenomena can trigger cascading failures that can occur before grid operators can respond. To bridge this gap, this paper first introduces the generalized induced failure N-k (N-k.k') problem, where an initial set of k failures induces an additional set of k' failures. In this context, we propose a disjoint optimization approach that decouples the computation of cascading failures from the interdiction model. This system response model, furthermore, is treated as a black box module without compromising the underlying performance guarantees of the interdiction model. Extensive experiments on several benchmark power grid case studies showcase the efficacy and scalability of the approach.

Compute cost in automated essay scoring (AES) is usually treated as a technical efficiency issue rather than a measurement design concern. Yet deployment decisions require more than reporting latency, model size, or inference expense. In high-stakes settings, total burden includes validation, monitoring, adjudication, compliance, and governance; in instructional settings, pricing, licensing, and integration may dominate cost. This paper reframes AES as a cost-constrained measurement system and introduces the Cost-Aware Measurement Design (CAMD) Framework, a four-layer structure consisting of metering, allocation, unit economics, and decision, for translating computational and workflow resource records into decision-relevant cost constructs. Direct cost is specified as monetary outlay plus attributable labor and time, and two criteria, design attribution and auditability, determine what the cost boundary excludes. Burdens that satisfy neither criterion, such as institutional capability and governance posture, are represented as constraints on which designs are admissible rather than priced. The paper defines cost per defensible score (CPDS) and argues that AES systems should be compared by quality-cost sufficiency: minimizing cost among defensible scoring designs that satisfy measurement requirements. Scoring volume conditions all four layers, so cost per defensible score is treated as a function of scale rather than as a single transferable number. CAMD provides a portable basis for representing, comparing, and evaluating cost across AES studies and deployment contexts.

Understanding neural scaling laws is increasingly important for theory and practice, yet first-principles explanations remain limited even in simple settings. In this work, we study capacity-limited scaling in a bottlenecked bigram model with power-law (Zipf) token statistics, viewed as an analytically tractable toy model of a large language model. Fixed random embeddings serve as a proxy for representational capacity and induce an explicit bottleneck of width, allowing us to derive scaling laws for prediction regret as a function of model width, vocabulary size, and the Zipf exponent. Our results complement recent theoretical accounts based on optimization dynamics by isolating scaling due purely to representational constraints. In the appropriate regime, the theory recovers the functional form of classical neural scaling laws and suggests that heavy-tailed language statistics may underlie both optimization- and capacity-driven scaling. Experiments natural language datasets (WikiText2, OpenWebText) closely match the predicted curves.

Modern machine learning models increasingly operate on variable-size structured data, such as sets and graphs. A central requirement is size generalization: the ability of a model trained on inputs of one size to generalize to inputs of a different, potentially much larger, size. Classical statistical learning theory, however, primarily considers fixed input domains, and its guarantees often deteriorate with input dimension. Although extensive work has established when models on sets and graphs can be transferred across input sizes, such results do not characterize the statistical cost of doing so: the sample complexity of size generalization, and its relationship to classical in-domain learning, remain largely unknown. We develop a statistical learning theory of size generalization and establish a sharp exponential separation in sample complexity. Letting $n$ denote the number of training samples, we show that polynomial source decay is insufficient for efficient size generalization: while it yields error $n^{-\Theta(1)}$ and hence polynomial sample complexity for in-domain learning, size-generalization error decays only as $(\log n)^{-\Theta(1)}$, yielding an exponential separation. Thus, polynomial in-domain sample complexity can become exponential across sizes. We then identify an exponential source condition under which size generalization achieves error $n^{-\Theta(1)}$, recovering polynomial sample complexity. These guarantees are independent of input dimension and depend only on the number of training samples, yielding dimension-free scaling laws. We also provide matching minimax lower bounds, establishing optimality. To our knowledge, this work provides the first dimension-free minimax theory of size generalization and a rigorous characterization of when, and at what statistical cost, learning can transfer across input sizes.

High-performance camera-based perception pipelines must sustain per-frame real-time budgets on resource-constrained embedded GPU platforms. Availability attacks such as SlowTrack inflate the persistent state of multi-object trackers: data-association workload accumulates across frames, so per-frame latency increases and drives the pipeline past its deadline. Detector- and NMS-layer defenses do not directly control this persistent tracker-side resource. We present TrackShield, a lightweight runtime resource-management layer that protects the tracker at the point where attack-induced workload accumulates. TrackShield combines low-overhead anomaly monitoring with staged birth capping, pruning of young or weakly matched tracks, and per-stage association-input admission control. Its hard admission guard provides a deterministic bound on admitted association workload, independently of the heuristic monitor. On a Jetson AGX Orin, TrackShield reduces attacked tracking-stage latency from $139.90$ to $6.27$\,ms and whole-pipeline P99 latency from $190.9$ to $48.8$\,ms, at a benign MOTA cost of $1.06$ points. As an availability defense, TrackShield focuses on keeping the tracker responsive under attack instead of recovering the full tracking accuracy degraded by adversarial detections.

Lane-based RISC-V vector processors exploit data parallelism well but pay a recurring cost for two common patterns: reductions, which require a separate tree pass, and deep dependent expression chains, which cannot be fused and are issued as serial vector instructions. We present TALU, a 5-stage RISC-V core whose execute stage is an NxN grid of registered ALU tiles instead of a single ALU. The grid is a reconfigurable, feed-forward spatial fabric: tiles route operands only from non-backward neighbours, and added diagonal forwarding lets a reduction collapse spatially. A minimal custom extension configures the per-tile operation, routing, and per-tile constants, and streams operands through the array. On a single fabric, TALU runs ordinary scalar code, single deep-path expressions, and multi-lane vector kernels. On the benchmark suite of the RISC-V2 vector processor, TALU is 1.8-5.5x faster than RISC-V2 and 4.5-15.5x faster than Vicuna at equal lane count, with a reduction cost (dot{\div}add) of 1.05x versus 1.96x and 1.26x for the vector cores. A retimed implementation places and routes at 1.03 GHz in 45 nm in 0.30 mm2, half the area of RISC-V2.

Atmospheric turbulence creates complex spatial and temporal distortions that significantly reduce image quality. This work compares two deep learning methods for turbulence reduction: Generative Adversarial Networks (GANs) and Transformer-based models. Both methods are trained and tested on the SOTIS dataset, which includes multi-frame sequences under different turbulence conditions with ground truth references. Our results show an important finding: GAN-based models perform better in severely degraded, low-quality situations, while Transformer-based models achieve better results in moderate to high-quality settings, especially in correcting geometric distortions. We also enhance the Turbulence Mitigation Transformer (TMT) by introducing a refined training strategy, an augmented input design that includes a temporally averaged frame as prior structural information, and a modified loss function that incorporates the Structural Similarity Index (SSIM). These improvements lead to consistent performance gains across various turbulence conditions. Overall, this study provides an in-depth evaluation of deep learning methods for turbulence reduction and offers practical guidance for selecting models in different imaging environments.

On-policy distillation provides dense supervision on the student's own samples, but its targets still come from the teacher. Verifier reinforcement learning (RL) instead uses an outcome signal, but it provides no direct constraint on capabilities outside the verifier. Combining the two is difficult for diffusion language models (dLLMs), because bidirectional denoising makes the policy likelihoods required by ratio-based RL unreliable. We develop a likelihood-free pipeline for a mixture-of-experts dLLM. Each rollout group is routed by verifier outcome. Mixed groups use ratio-free, group-centred REINFORCE. If every student rollout fails but a teacher succeeds, the model receives dense KL supervision on the student's denoising canvases. All other groups are skipped. A KL term anchored to the distilled checkpoint limits changes outside the verifier reward. The tool-calling teacher succeeds on only 18.6% of the training prompts. At 24 denoising steps, distillation exceeds the teacher by 6.77 percentage points, and routing improves the distilled checkpoint by a further 18.31 points. Routing lowers reasoning-trace accuracy by 0.71 points but raises direct-answer accuracy on the same items by 13.21 points. The anchored arm restores trace accuracy. Its tool-calling score differs by -0.16 points, with a confidence interval that includes zero, although this comparison also removes a small fallback distillation route. A domain-conditional recipe retains a 16.37-point tool-calling gain; a second seed gives 18.23 points and reproduces both preservation results, but the math result is seed-dependent. We pre-registered every arm and evaluated each comparison on paired items in a single sweep on one machine.

The ecological insurance theory posits that diverse forests are more resilient and better able to adapt to changing environmental conditions. Increasing drought and heat stress in Central Europe promotes mixed forests for maintaining ecosystem services. However, a diverse species distribution complicates remote sensing-based species mapping, as distinguishing pure from mixed species compositions at the pixel level remains challenging. We evaluated the potential of recurrent neural networks (RNNs) for tree-species classification from Sentinel-2 time series and investigated the spectral-temporal features underlying class discrimination, with particular emphasis on separating pure from mixed Norway spruce pixels. We compared RNN classification models with Random Forest (RF) and Extreme Gradient Boosting (XGBoost) and used Integrated Gradients (IG) to quantify feature attributions across spectral bands and time. The best-performing RNN achieved a Macro-F1 score of 0.851 on the test data, outperforming XGBoost by 0.012 and Random Forest by 0.103. Beyond classification performance, the attribution analysis revealed distinct spectral-temporal patterns associated with the separation of pure and mixed Norway spruce pixels. Attribution values increased particularly after approximately day of year (DOY) 125, while B11 showed the strongest contribution during periods of increasing spectral divergence between the two classes. These findings show that recurrent models can provide competitive classification performance while also revealing the phenological periods and spectral regions that are most relevant for distinguishing closely related forest compositions. Combining time-series classification with explainability therefore provides additional insight into the spectral-temporal mechanisms underlying tree-species discrimination and may support more detailed mapping of mixed forest structures.

Benchmarks shape how model capabilities are measured and interpreted, yet their scores can obscure whose conventions, contexts, and judgments define successful performance. We introduce a user-centric audit framework that distinguishes users who contribute benchmark content from those whose needs evaluation results are intended to represent. The framework operationalizes seven dimensions across benchmark content, metadata, documentation, and scoring procedures. Across 16 resources, we find that American-centered conventions account for 62.4-94.4% of matched signals in 11 text benchmarks, while geographic visibility in five image-text resources is concentrated and strongly associated with national GDP share (mean Pearson r = 0.780; Spearman p = 0.778). Documentation reveals a systematic source-target gap: all 16 resources operationalize source-user context, but none operationalize target-user context. Paired scoring experiments further show that meaning-preserving answer variants reduce DROP exact-match accuracy from 100.0% to 33.3%, while TruthfulQA exhibits metric-dependent reductions in reference similarity. Our framework and documentation checklist make user assumptions auditable, clarifying what benchmark scores represent and where evaluation claims are supported.

Online memes are often carriers of unsafe content such as disinformation and propaganda. Their rapid evolution in response to news events and strong dependence on contextual information make it difficult for platforms and moderators to detect and moderate them effectively. Moreover, existing moderation studies typically classify memes as safe or unsafe based on individual images, lacking the ability to perform broader provenance analysis, i.e., where a meme originates, how it spreads, what cultural symbolism it represents, whether it poses potential risks to society, etc. To bridge this gap, we introduce MemeTrace, a multimodal LLM-based agent that automatically traces the provenance and potential risks of emerging memes. We evaluate MemeTrace on Meme100, a dataset of 100 real-world memes, using both human annotation and automatic metrics. The results demonstrate that MemeTrace substantially outperforms the VLM baseline. More importantly, MemeTrace achieves higher accuracy in detecting unsafe memes, especially those involving disinformation and sensitive content, and remains robust under visual perturbations.

Existing illumination estimation methods primarily recover lighting distributions for rendering, but typically leave the contributions of individual physical light sources at specific scene locations implicit. We address this gap by leveraging the scene understanding capabilities of vision-language models (VLMs) to connect local illumination with its physical sources. We introduce LAMP, a framework for point-conditioned, source-aware illumination reasoning from a single RGB image with marked query points. LAMP constructs a shared inventory of active light sources, ranks contributing sources for each query, and predicts their 3D query-to-source directions. We supervise these relationships using procedurally controlled indoor scenes, reusing independently rendered source responses to derive contribution targets across lighting configurations and obtaining exact directions from source and query positions. On the Infinigen test set, fine-tuning Qwen3-VL-8B increases dominant-source 3D PCK@0.5m from 0% to 70.6%, with 94.5% of directions within 12.5 degrees of ground truth among correctly localized sources. Combined with a lightweight network predicting dense RGB irradiance, LAMP achieves probe-rendering quality comparable to prior lighting-estimation methods while additionally localizing individual physical sources in 3D and explicitly attributing local illumination to them.

Agent skills are Markdown documents that a tool-using agent reads as instructions and then acts on with the privileges of the session that loaded them. Installing one grants authority (the standing to direct the agent, not new privilege), and the size of that grant is what we measure. We present a two-phase attack-surface profiling methodology that combines whole-bundle static analysis and dependency provenance checking with instrumented sandbox execution and separate model-based assessment of the resulting traces. We apply the methodology to 6,309 listings from two marketplaces, selecting listings for execution without conditioning that decision on a prior threat or risk signal. Half the corpus carries evidence of shell-execution capability; the corresponding shares are 35% for filesystem read, 31% each for filesystem write and network egress, and 13% for credential access; seven listings in ten carry at least one elevated capability. That is the surface a user accepts at install time, and almost none of it is visible then. A permission model exists but is close to inert. Authors may list the tools a skill pre-approves in its own frontmatter, which removes confirmation prompts rather than restricting what the skill can reach; of the 6,061 listings whose frontmatter parses, 770 declare anything at all, the vocabulary cannot express credential or environment access, and nothing checks what it does express against the bundle. Declaring is also not declaring accurately: 59% of declaring listings understate what their bundle can do. ClawHub declares roughly twice as often as skillsmp, yet under-declares on 88% of its declarations against 31%. Triage at the marketplace boundary therefore cannot rest on what a listing declares. It must rest on what the bundle can do. These results position capability-oriented attack-surface profiling as a necessary complement to maliciousness detection.

Robots following persistent language instructions must decide which cached room judgments remain valid as their environment changes. We present MIND (Maintenance via INstruction-conditioned Dependencies), a framework for maintaining target-room decisions without re-evaluating every room after each event. A language model compiles each instruction into dependencies for semantic qualification and target selection. The controller intersects semantic dependencies with event footprints, refreshes affected judgments, and updates reachability and selection separately. For fixed deterministic evaluation, planning, and selection, conservative dependency and footprint coverage preserves the decisions of full recomputation. On 240 five-event ProcTHOR sequences with an exact benchmark predicate, MIND reduces mean semantic evaluations from 30.6 to 11.2 while preserving all 1,200 post-event decisions. Frozen-evaluator experiments with Phi-4 and Qwen3-14B also preserve all 1,200 decisions, without improving the evaluators' underlying task accuracy. In 32 closed-loop AI2-THOR episodes using exact-predicate qualification and metadata-derived footprints, MIND matches the final target decision of full recomputation in every episode. Unitree G1 experiments demonstrate target reuse and reselection across room transitions without additional semantic evaluations under MIND, and abstention after a barrier and an object removal at 4 evaluations against 15 to 18. Additional materials are available at https://mind-icra27.github.io/.

Memory for long-horizon video generation must satisfy two competing requirements: it must remain bounded so generation can continue indefinitely, while still recovering what was observed at a place when the camera returns, regardless of how long ago it was seen. Reconstruction-based memories provide spatial recall by caching and reprojecting explicit geometry, but require growing storage and depend on estimated 3D structure. Recurrent memories keep a constant-size state, but retention is governed by temporal recurrence, causing old observations to fade with time. We introduce SMART, a slot memory addressed by region tokens, which combines bounded storage with spatially addressed recall. SMART stores content in a constant-size slot recurrence whose writes are addressed by the observing camera and whose reads are queried by the rendering camera, without constructing an explicit 3D representation or maintaining a growing cache. Consequently, recall is conditioned on inter-view geometry rather than on recency. Each read additionally provides a confidence estimate that conditions a chunk-causal diffusion transformer, allowing it to selectively trust retrieved content while relying on its generative prior elsewhere. On long trajectories in Memory Maze and Minecraft, SMART maintains high-fidelity recall far beyond its training horizon while using constant memory and per-step cost.

Physics-informed neural networks provide a flexible framework for learning parametric PDE solution families without labeled data, yet preserving accurate derivatives across heterogeneous regimes remains challenging. Structured neural solvers offer compact representations, but restrictive conditional parameterizations can limit adaptation across the solution family. We introduce BeTINN, a structured conditional architecture extending time-induced modulation through multi-directional time-parameter conditioning and a joint residual pathway. We evaluate BeTINN on the Heston stochastic-volatility PDE, a demanding parametric problem with mixed derivatives and economically relevant first- and second-order sensitivities. BeTINN improves joint accuracy across normalized option value and key sensitivity metrics relative to our baselines. Controlled ablations characterize the roles of its architectural components. We further find that validation physics loss is informative but does not fully determine differential fidelity, and that accurate solution values can coexist with localized derivative errors. These results support structured conditional adaptation and motivate derivative-aware evaluation when sensitivities are downstream quantities of interest.

Fairness audits with several sensitive attributes are often performed one coordinate at a time. This can miss purely intersectional disparities: a score may be independent of each sensitive coordinate while depending on their joint profile. We study distributional demographic parity for a real-valued score using one-dimensional Wasserstein geometry. We distinguish two complementary constructions. Conditional and sequential audits characterize joint parity through their zero sets, but their positive values do not generally decompose a fixed joint Wasserstein disparity. By contrast, full-cell conditional quantile functions define a Hilbert-valued random variable whose variance is the joint disparity. Conditional expectations of this quantile process induce a monotone information game: every reveal order yields an exact martingale decomposition of the same joint disparity, while Shapley values provide an order-free attribution across sensitive attributes. Under independent sensitive attributes, the game reduces to a Hoeffding-ANOVA decomposition and Shapley splits each interaction energy equally among its members; under dependence, orthogonality is lost and explicit formulas reveal suppression and redundancy effects that marginal audits can hide. We give fixed-grid estimators and permutation calibration, and illustrate the framework on synthetic examples and ACSIncome. The resulting framework separates three questions that are often conflated: whether joint distributional parity holds, how a fixed joint disparity decomposes along an information path, and how that disparity can be attributed across sensitive coordinates without privileging an ordering.

Reduced-order models (ROMs) make model predictive control (MPC) of high-dimensional nonlinear systems tractable, but the resulting modeling error can lead to constraint violations on the real system. We introduce CoRE-MPC, a robust MPC formulation that plans with a nonlinear ROM and guarantees recursive feasibility and constraint satisfaction for the full-order system. It can be applied with a broad class of nonlinear model reduction techniques and the resulting MPC does not require evaluating the full-order model online, thus ensuring computational efficiency. The key insight is in how the ROM affects the dynamics of the prediction error. We account for this with two error radii that are propagated alongside the ROM and provide a guaranteed bound on the prediction error. The approach is validated in simulation on an 80-state oscillator chain, a Burgers equation, and a 9,768-state finite-element soft robot, where a two-state ROM with data-driven bounds keeps all 1,628 nodes clear of a wall at a median solve time below 1 ms for a 20 ms control period. Across these examples, CoRE-MPC is 13 to 61 times faster than full-order robust MPC at only 27 to 35% higher closed-loop cost, whereas existing certified schemes are either infeasible or at least twice as conservative. Code: [released upon publication].

Robotic disassembly of end-of-life (EoL) products requires robots to accommodate product variation, follow prerequisite sequences, and approach components precisely for contact-rich manipulation. Supervised fine-tuning (SFT) provides a practical path for adaptation, but limited and costly disassembly data cannot capture the distribution shifts arising from changing environments and product uncertainties. The policy's own rollouts record its approach attempts, and human corrections demonstrate recovery from its errors, but turning these experiences into better policy performance remains challenging. We introduce PACER, a Process-Aware Correction and Evidence Reweighting framework that learns from these experiences through SFT, using progress toward the target and each trace's source and outcome to weight valid action chunks in the native supervised action loss. PACER trains one policy per weighting setting and selects one of the resulting policies through held-out validation. We evaluate PACER on UR5e motherboard disassembly across five component classes, where it achieves 87/100 successful approach-and-handoff trials on the held-out benchmark, compared with 54/100 for vanilla post-SFT.

Transformers are intuitively considered to be efficient at copying from their inputs. We actually show that the ability of LLMs to copy in context can be improved through explicit modeling. We propose Copy-Paste Language Models (CPLMs), which add a lightweight Copy-Paste head to the conventional language modeling (LM) head. CPLMs improve training efficiency, downstream performance, and offer new operating points in long-context and retrieval setups. We validate our approach by pretraining models from 135M to 1.7B parameters, showing improvements across all scales and backbones with negligible training and inference overheads. We explain the success of CPLMs by showing that they partially mitigate the LM head gradient bottleneck through an alternative gradient path.

Equivariant structure-prediction pipelines alternate coordinate updates with the construction of a k-nearest-neighbour (kNN) graph whose membership is a discrete, piecewise-constant function of the coordinates. Rebuilding that graph at every block boundary (PB) is the expensive part of an epoch, whereas reusing one frozen graph (F) is fast but leaves a topology gap. We study a two-pass design in which the cheap pass records a small amount of state (margins, kNN radii, measured per-boundary displacements) and a second pass compensates conservatively. We formulate the choice of the graph-construction basis as a damped fixed-point iteration whose unique fixed point is exactly PB, and derive (i) exact non-flip certificates with the two-sided constant $m_i/4$ and one-sided constant $m_i/2$, (ii) the minimum-variance (BLUE) replay gain $\gamma^*$ and the resulting feasibility crossover, (iii) an $\ell_\infty$ trust-region projection that preserves convergence, and (iv) an analytic near-tie spacing law $E[m] \approx r_k/(3k)$ that matches measurements to 1%. Three compensation levels are implemented in C++ (L1: frozen set with live geometry; L2: margin-gated local repair; L3: damped, projected basis) and validated under a matched-conditions protocol (identical seed, MSA depth and trajectory family, only the topology mode differs). L2 reproduces the PB edge set exactly ($D = 0.0000$ at every boundary), L1 stays within 0.6% and L3 interpolates monotonically: L2 reproduces the PB edge set exactly when the trust region is removed, and L3 interpolates between the frozen and the PB topology. The recorded trajectory itself yields a measured $1.5\times$ Pass-1 and $1.76\times$ epoch speed-up with no measurable loss gap. An open-source implementation of the runtime and of all the experiments reported here is available at https://github.com/AlbertHuang96/ProteinPredictMachineLearning.

Financial disclosures express the underlying relationships of double-entry accounting through unstructured text. While financial LLM benchmarks predominantly emphasize question answering and classification, schema-guided evidence discovery remains relatively under-explored. We investigate whether open-weight LLMs can extract and classify evidence corresponding to predefined double-entry accounting relationships underlying the text. Using a financial accounting graph comprising 11 nodes (accounts) and 11 edges (transaction types), we evaluate mid-sized models from 5 open-weight families across 1,300 SEC 10-K Item 7 filings spanning 26 years, comparing their extraction performance against three proprietary models. Results reveal substantial differences in extraction volume, edge distributions, and cross-model agreement. Building on this study, we introduce a human-annotated relation classification benchmark comprising 483 passages across three financial accounting categories, evaluating 34 open-weight and 10 proprietary models. The strongest open-weight and proprietary models achieve 84.1% and 92.8% accuracy, respectively, with considerable variation across accounting categories. Our findings highlight the importance of evaluating evidence discovery and relation understanding separately, while investigating locally deployable models for scalable, privacy-sensitive financial analysis.

In this note we explore the following simple question: How does the classic gambler's ruin phenomenon change if we allow the gambler financial leverage? We show that, with a moderate amount of leverage, it is possible for an (idealized) gambler to sustain seemingly improbable profits over a remarkably long time horizon without ever going bankrupt, even if the gambler makes only fair bets. With a leverage ratio of $L = 10$ (less than what Lehman Brothers Holdings had from 1995–2008), the probability of multiplying one's money by $1000\times$ without going bankrupt is roughly $50\%$. More generally, given $L$-fold leverage, and given the goal of multiplying one's money by a factor of $r > 1$, the optimal gambling strategy has probability roughly $r^{-1/L}$ of success. The results in this note are designed as part of the course development for the CMU randomized algorithms course.

When writing code, the same computation can be expressed in different programming languages and through different implementation strategies. We show that the activations of LLMs capture this shared structure in reusable and composable abstract representations when they generate code. These representations emerge in the middle layers, are reusable across programming languages, and afford the model an impressive amount of generalization ability when generating code under various interventions. Using activation steering, we show that function, programming language, and implementation strategy can be independently manipulated. The independent interventions on language and function can be flexibly composed to control code generation, even in cases when the prompt contains no information about language and function. We additionally find that finetuning on examples of arbitrarily defined functions in one programming language transfers to other languages, consistent with the models reusing abstract features of code generation that are shared across different examples. Finally, to study how such abstract representations come to be, we implement a minimal model of shared, reusable, and composable structure in a synthetic setting, and show that transformers trained on such data learn representations that reuse these parts flexibly across contexts.

Automatic stuttering systems primarily detect dysfluencies after they have occurred, limiting their use for anticipatory intervention. We address this gap by formulating stuttering prediction as continuous causal time-to-event forecasting, where a model estimates the risk of an upcoming stuttering onset every 100 ms using only preceding speech. To support this task, we augment FluencyBank Timestamped for 36 speakers with manually reviewed repetitions, prolongations, and blocks, yielding 2,872 forecasting events. We then model the temporal evolution of frozen Whisper-large-v3 speech representations using a compact causal Conformer with 3.67M trainable parameters. The proposed approach achieves an AUROC of 0.63. Temporal analysis further shows that explicitly modeling preceding speech states outperforms simply increasing the static acoustic window, while a discrete-time survival formulation provides coherent forecasts from 300 ms to 1.5 s using a single model. These results provide initial evidence that stuttering onset is preceded by measurable temporal information in speech and establish a framework for future anticipatory stuttering systems.

Vision-language models (VLMs) can reason about visual planning tasks, but querying them for every instance incurs repeated inference cost and latency. Generating reusable planning programs offers an alternative, but constructing reliable planning systems for new problem domains without extensive supervision remains challenging. We present a training-free framework that constructs reusable perception and planning programs from a domain specification and a single practice scene without a paired solution. The central challenge is obtaining useful feedback from the limited input. Our framework addresses it by leveraging VLM's ability to provide a renderer and a dynamic model based on its scene understanding. Rendering detected states enables comparison with the practice scene, while checking simulated action outcomes against the domain specification helps expose errors in rule implementation. Together, these checks guide refinement of the grounding program and planner without requiring solution labels. After refinement and component selection, the programs are frozen and solve new instances without further VLM calls. The same framework constructs systems across domains without manual pipeline redesign. Across fifteen domain variants spanning 2D and 3D scenes with discrete and continuous state, our systems achieve 99.3% average planning success, compared with 80.9% for the strongest domain-level program-generation baseline and 90.7% for the strongest instance-level baseline. Cost analyses show a favorable success-cost trade-off, and ablations confirm the contributions of the executable domain model, execution feedback, and component selection.

Large language models impose substantial memory demands and arithmetic costs during inference. Dense-to-mixture-of-experts (MoE) conversion is a model optimization technique that aims to reduce these arithmetic costs. Dense-to-MoE conversion partitions pretrained feed-forward network (FFN) neurons into experts and selects a subset for each token. Leading methods fall into two categories: analytic methods produce a converted model quickly using heuristics, while learning methods produce a model much more slowly by optimizing via gradients. Both approaches, however, fix the same active expert count, or "utilization," at every layer. We introduce Spectral-MoE, which learns to allocate a utilization budget heterogeneously across layers. Its name draws a metaphor between separating a dense FFN into experts and separating a light beam into its spectrum. Spectral-MoE consists of two distinct approaches: a static approach that uses a single per-layer scalar which is shared across tokens, and a dynamic approach which emits a per-token schedule. We demonstrate that both allocators improve performance over uniform on both the analytic-based and learning-based SOTA dense-to-MoE methods. At 25% FFN utilization, dynamic allocation reduces WikiText-2 perplexity relative to uniform by 1.04, 2.96, and 2.71 points on the learned base for Llama-2-7B, Llama-3-8B, and Qwen2.5-7B, respectively. The corresponding reductions on the analytic base are 2.45, 6.26, and 2.17 points. On the learned base, dynamic allocation improves six-task mean accuracy over uniform by an average of 1.5, 4.2, and 3.6 percentage points across the three models at FFN utilizations of 12.5%, 25%, and 50%, respectively. These results establish per-layer expert utilization as a novel optimization axis in dense-to-MoE conversion.

Transformer language models develop attention heads that specialize in syntactic relations, for example, attending from verbs to their subjects. Do these heads play a causal role in the model's syntactic predictions? Evidence for such a role is mixed to negative: even when we force these heads to attend to a distractor noun other than the subject, the model's syntactic predictions are still more consistent with the subject than with the distractor. Interventions on specialized syntax heads had a significantly stronger effect on syntactic predictions than interventions on random heads, but only in one of the two measures we propose.