Science & Mathematics

World models let agents compare imagined outcomes before acting, but an agent must still decide which alternatives to explore and when a proposed plan should replace its learned policy. We introduce SDI (Sibling Disagreement in Imagination), which connects these decisions through a shared actor-based reference. SDI compares action sequences imagined from the same model state, using repeated rollouts of the actor's own sequence to measure variation under fixed behavior. This reference scales both the spread across candidate returns, which provides an exploration signal, and a proposed plan's predicted advantage, which governs action acceptance. The agent executes a planned action when its advantage exceeds the reference variability by a fixed factor; otherwise, it follows the actor. SDI reuses the base agent's world model, actor, and critic without requiring an additional disagreement ensemble. Exploration and planning guide training data collection, while evaluation uses the learned actor alone. Across Meta-World, Adroit, and DM-Control, SDI achieves a higher observed task-average score than every compared baseline in each suite and matches or exceeds the strongest baseline's performance on 15 of 18 tasks. Component ablations and robustness analyses further examine the roles of exploration, planning exposure, and action acceptance. Paired evaluations from identical simulator states show that the gate improves both recovery from actor failures and preservation of successful behavior relative to ungated planning, while substantially reducing return degradation. Together, these results support comparing imagined alternatives against a shared actor reference as a unified approach to exploring and planning with world models.

Aggressive low-state weight quantization can substantially reduce neural-network representational flexibility, making predictive performance difficult to preserve when only a few discrete weight states are available. We investigate this regime for Transformer-based electricity-load forecasting using 3-, 5-, 7-, and 16-state weight quantization across feed-forward-only and attention-plus-feed-forward quantization. Two tasks predict either the next hour or the next 24 hours from the preceding 48 hours across two co-primary electricity-consumption datasets. Direct post-training quantization (PTQ) increased mean normalized root-mean-square error (NRMSE) relative to full precision by 6.25 and 7.04 percentage points in Datasets A and B, whereas quantization-aware fine-tuning (QAT-FT) recovered approximately 90.2% and 94.5% of these penalties. Quantization-aware training from random initialization (Scratch-QAT) also produced competitive low-state solutions, with particularly favorable behavior at three states. We further introduce accumulated-update thresholding, which accumulates optimizer-proposed weight changes before commitment. Relative to paired Scratch-QAT controls, this reduced committed shadow-weight updates by approximately 99.7% and discrete state transitions by approximately 77--80%, while mean NRMSE changed by only +0.042 and +0.197 percentage points. These results show that useful Transformer forecasting can be retained under highly discrete weight representations while strongly suppressing target-state activity, supporting a prospective route toward compact, low-activity edge forecasting using multilevel synaptic devices.

Training tool-using language models requires costly interaction, yet identical rewards within a rollout group provide no group-relative learning signal. We use reward disagreement to guide harness evaluation and the allocation of rollouts for policy training in Model Context Protocol (MCP) environments. We first introduce Discord, which tests harness edits using exact paired inference and decision-preserving early stopping: only disagreeing paired outcomes favor one harness over another. This insight motivates prospective allocation in BRACE (Bank-calibrated Rollout Allocation under Correlated Episodes). BRACE combines historical task evidence, uncertainty in task success rates, and a correlated reward model to forecast mixed-reward groups before allocating generation budget. Controlled harness experiments favor Discord, while live MCP experiments expose a retention-cost trade-off. At three Qwen3-VL model scales, BRACE exceeds the strongest plotted baseline peaks with fewer new training rollouts; historical-bank collection is counted separately. It also leads the compared methods on BFCL and NESTFUL without further training. Ablations probe the bank, estimator, and allocation score; prospective evaluation checks forecast calibration. These results support forecasting reward contrast before generation.

Evaluations of large language model (LLM) agents report that pricing agents collude at others' expense. Yet collusion is defined by enforcement: a cartel, in which competitors share monopoly profits, holds only if a member who cheats is punished. We introduce an audit that tests enforcement by forcing a member to cheat in counterfactual copies of a market, and use it to study what reinforcement learning (RL) teaches agents in pricing and trading. Small LLMs trained by RL learn what their training rivals reward, but only what they readily try. In self-play, they learn fixed rules ignoring their rivals; against a punishing rival, to be deterred; against a cheating rival, to match its cuts, never the costly price war that would deter it, which they rarely try. A war forced into training is learned, but mostly only where it was forced. Against their own copies, deterred and matching agents can earn the same cartel profits, but only matching ones can enforce them, and then thinly: as a by-product and only against large cuts. Q-learners enforce collusion only in pricing. Prompted LLMs reach cartel-like outcomes mostly through rules that ignore the market. Outcome measures cannot tell these cases apart; the audit can.

Flow-matching policies adjust inference cost through the number of solver steps, but reinforcement learning at a single fixed solver budget can hide degradation at cheaper deployment budgets and even erase useful one-step behavior present in a pretrained policy. We introduce Action-Ratio-Clip Flow (ArcFlow), a deployment-aware algorithm for training one flow policy across a specified range of solver budgets. ArcFlow combines action-aligned supervision, budget exposure, and worst-budget checkpoint selection. For reconstruction objectives, its bridge aggregates temporal probes into one surrogate ratio and one PPO-style clip per action, with a boundary residual exactly equal to deployed one-step reconstruction error. Cycling exposes the shared field to the deployment range, and selection optimizes its performance floor. For native velocity objectives, multi-budget endpoint supervision applies the same principle without replacing the native update. Both realizations preserve the architecture and deployment-time cost. Across DM Control, manipulation, and locomotion, ArcFlow attains the highest mean one-step return among five flow-policy baselines on 25 of 30 tasks under each method's reported training settings. Component studies identify complementary improvements from boundary supervision, the fixed-grid bridge, budget exposure, and deployment selection. These results establish deployment-budget-aware training as an effective route to strong low-step control with one policy serving multiple inference costs.

Target-based extrinsic calibration estimates the rigid transform between sensors from a few observations of a known target, and its result is commonly assessed by the root-mean-square error of the fitted residuals. However, this scalar error is computed on the same observations and therefore cannot indicate whether to accept the result, re-extract, recollect or add observations. This paper presents VeriCalib, which formulates extrinsic calibration as a verifiable estimation problem and reports a calibration state. The key insight is a two-level extrinsic compensation mechanism, in which the extrinsic compensates the rigid part of each observation error at the residual level, and rotation and translation compensate each other at the parameter level. Consequently, the fitted residual can reveal neither an inconsistency between the observations and the model nor whether the estimated extrinsic is reliable in every direction. Accordingly, we assess the calibration through a structured covariance that jointly propagates the uncertainty of the camera and LiDAR board features with a known-geometry prior, and estimate its scale as a variance component from the residuals. On this covariance, we decouple rotation and translation to resolve the precision along each direction and guide data collection toward the weakest one. Since this precision presumes consistent observations, we further develop within-view and cross-view checks that examine the internal residual and the rigid discrepancy between observations, respectively, with thresholds that hold their significance level under an unknown noise level. We also extend the minimal detectable bias to the extrinsic to bound the smallest detectable fault along each direction. These criteria form a calibration state that indicates whether to accept the result or how to collect the data...

Software supply chain attacks continue to be on the rise, exploiting the trust that helps open source ecosystems thrive. Recent incidents like the Shai-Hulud worm and the Axios compromise demonstrate that a single malicious package can have wide reaching and devastating impacts overnight. These high profile attacks can be traced back to the same culprit: package registries. Although they are a useful convenience to pull prebuilt packages from, is such a commonly exploited component of the software supply chain worth the risk? In this paper, we investigate whether removing the registry from the developer workflow by building packages directly from source can serve as a malware prevention strategy - effectively removing an entire attack surface. Our hypothesis is that malicious packages avoid attributing a source to minimize visibility, while benign packages seek exposure and therefore link to their repositories. We evaluate this through three research questions: how effective is building from source at avoiding malware, how feasible is it to build from source in the current state of the supply chain, and how the building from source strategy compares to state-of-the-art security scanners. To do so, we (1) construct an aggregate dataset of 47,066 malicious packages from four academic and one industry malicious datasets source across npm and PyPI ecosystems, (2) build a snapshot index of all 3.9 million npm and 764 thousand PyPI packages, and (3) analyze all malicious samples alongside a statistically significant sample of 16,569 packages from each registry, totaling 33,138 packages. We find that simply checking for a source has a success rate between 87.7 and 99.9 percent of preventing malware, depending on the level of source vetting. Consequently, there is a significant lack of transparency within benign packages, with only 47.4 percent attributing a source an end user could build from.

We introduce CALLIOPE, a top-down framework for discovering topic hierarchies in large corpora of unstructured data (e.g., LLM conversations), while supporting differential privacy (DP). Unlike traditional and even recent LLM-augmented approaches that ultimately rely on clustering vector embeddings, our framework leverages LLM capabilities directly for the entire pipeline. Experiments demonstrate that this approach provides better utility over existing (embedding-based) methods in the DP setting, especially in the high privacy regime. We evaluate the quality of CALLIOPE outputs using a variety of traditionally studied clustering metrics, as well as some novel utility metrics that we introduce.

Iterative ML agents develop models by editing code, running experiments, reading a validation score, and choosing what to try next. In current harnesses this loop is served by an exact validation score returned after every attempt, which leaves the validation split open to adaptive selection. In this work, we introduce an explicit feedback contract for ML-agent harnesses, which specifies what an agent may observe, when iteration should stop, and when the test split may be touched. The contract releases a validation score only when a submission clears a noise threshold computed from the paired standard error, so no margin has to be tuned. We ran experiments on click-through-rate prediction, learning to rank, and a control task with no signal to find. The control is the sharpest case, because the true gain there is zero by construction: an agent reading exact scores ended by reporting a gain of 0.0157, while under the contract it reported 0.0033. On both real tasks the contract narrows the validation-test gap, and test performance is slightly higher rather than lower. The largest difference is in when the agents stop. Exact feedback used its entire budget of thirty submissions on two of the three tasks, including the one where nothing can be learned, while the contract halted after four to twelve. A run that exhausts its budget is usually read as diligence. It is better read as a symptom: when the feedback channel is unbounded, noise keeps supplying improvements to chase.

In case-based legal reasoning, legal professionals identify legally significant fact patterns in a new dispute and compare them with those of decided cases. In AI & Law, HYPO dimensions and CATO factors represent these patterns symbolically, organizing expert knowledge distilled from case narratives as explicit, structured, interpretable, and auditable legal concepts and relations that support case comparison, legal argument, and precedent-based reasoning. Large language models (LLMs) are increasingly used for legal tasks, yet their internal organization of legal knowledge remains difficult to trace. We investigate whether LLMs form internal representations corresponding to all 26 CATO factors. Two research questions examine linear decodability and sparse localization with sparse autoencoders. Experiments cover three open models, Qwen3-1.7B, Qwen3-8B, and Gemma-2 2B, supported by cross-model comparison and human audit of controlled natural-language passages. Results show that all 26 factors are linearly decodable in all three models. Single-feature sparse readouts cover the complete factor catalog, with best-layer test AUCs ranging from 0.86 to 1.00. Five selected features approach full-vector readout performance. We report 78 model-specific layer-feature coordinates, one for each factor in each model. Together, these results show that information corresponding to expert-defined legal factors can be read from the internal activations of LLMs processing legal narratives. By locating these readouts, our study connects legal language and the symbolic abstractions of HYPO and CATO to learned internal representations.

We revisit the problem of depth separation for neural networks, as studied by Eldan and Shamir, Daniely, and others. Specifically, we consider the natural setting of $L^2$ approximation of radial functions under the standard Gaussian measure in $d$ dimensions. We show that there is a radial function in $L^2$ under the standard Gaussian measure that can be approximated to a fixed sufficiently small $L^2$ error by a depth-3 ReLU neural network of polynomial size in $d$ but can only be approximated by depth-2 networks with subexponentially bounded weights if their size is at least exponential in $d$. For that, we prove a general result showing that if a radial function has substantial mass in the high frequencies of its Hermite expansion, then it cannot be well approximated by a depth-2 neural network with small width and bounded weights. All theorems, lemmas, propositions, and corollaries are verified in Lean~4.

Most unified multimodal models treat images and text asymmetrically: a pretrained language model generates text, an attached diffusion head generates images, and outputs that contain both are produced in a fixed order, so the first output cannot adapt to the second. We ask whether both modalities can instead be generated by one process that is trained on them jointly from scratch. Pretrained encoder spaces hold both modalities in one continuous form and carry knowledge learned at scale, and recent work finds them effective as generative targets for vision and language. We therefore model the joint distribution directly in these spaces. We introduce JUICE (Joint Unified multimodal modeling In Continuous Encoder spaces), a single flow matching Transformer over frozen image and text representations, in which each modality has its own noise level. Which modalities are clean and which are generated then decides the task: image generation, captioning, visual question answering, or joint generation. Controlled pretraining runs test each design choice. At equal guidance, generating text as continuous states leaves image generation intact (48.41 GenEval, against 47.86 for a model trained only for text-to-image generation), while captions lose some exact wording. Training on joint and conditional denoising together beats either alone, and semantic encoder spaces beat VAE latents and raw pixels as targets. At 2.1B parameters, with no language model initialization and at 224x224, JUICE reaches up to 0.88 on GenEval and 73.1 on VQAv2. By sampling both modalities from the joint distribution (coupled sampling), it can also draw a route on a map while writing out the matching moves. On mazes with several valid solutions, joint success rises from 54.6% with independent sampling to 93.7%, and no modality has to go first. On visual spatial planning, JUICE solves 99.8% of maps at its training sizes.

How much does an agent harness affect vulnerability discovery, and does its value depend on the model? We study 24 model-runner configurations using an open-source harness on a production codebase whose vulnerabilities were non-public at scan time. The pipeline combines static screening, LLM investigation, model self-revalidation, root-cause mapping, independent source review, and isolated proof-of-concept confirmation. Across 265 confirmed security root causes, the largest two-run configuration covers 60.4%, while 14.3% are found by exactly one model. No single configuration covers all observed findings. In a four-model context experiment, moving from a minimal prompt to an engineered prompt with a CodeQL-derived call graph raises criticality-weighted yield by 20-86% for smaller and mid-tier models, compared with 2% for the frontier model. This comparison changes both prompt and context. Separating those effects shows that the engineered prompt improves yield for all four models, while adding CodeQL context beyond that prompt improves yield only for the smallest model. A held-out CyberGym subset shows the same direction of target-recall improvement for the combined treatment. Frontier models achieve low false-report rates but do not consistently achieve the highest yield. Strong lower-cost models can deliver similar performance at an order of magnitude lower cost. These results support matching models, prompts, and context to the task, and combining configurations for broader coverage.

While statistical learning has made remarkable progress in recent decades, driven by large datasets and data-driven protocols that have transformed scientific practice, it still faces a key limitation: it struggles to generalize beyond the data it has seen, often failing under distribution shift. Causality offers a promising solution: causal models focus on the underlying mechanisms of change rather than surface-level associations, enabling robust reasoning under external interventions. This allows systems not only to learn from data, but also to reason about hypothetical scenarios, adapt to new environments, and inform better decisions. This thesis explores how these approaches can inform each other, in two parts. The first part examines algorithmic recourse, that is, identifying which changes an individual can make to overturn an unfavorable decision. We propose an alternative perspective to the established framework, arguing for a more realistic cost function, a reformulation of the problem in causal terms, and the inclusion of a temporal dimension in policy evaluation. The second part focuses on dynamical systems, extending the notions of causal interventions and effects to discrete and continuous stochastic processes. We show that the equilibria of these processes, and how they change following an intervention, can be linked to Structural Causal Models (SCMs), viewed as compressed representations of the underlying temporal dynamics. We argue for the relevance of such temporal causal models, since in real-world scenarios actions take time to produce their effects. The common goal of both parts is practical: the utility, robustness, and trustworthiness of these models depend critically on their ability to realistically represent the underlying phenomena.

Community Choice Aggregators (CCAs) play an important role in the energy transition by procuring renewable electricity for local communities while balancing affordability, resource adequacy, and clean-energy requirements. In California, these long-term procurement decisions are coordinated through integrated resource planning (IRP). However, conventional IRP frameworks typically represent wholesale electricity prices deterministically, even though price uncertainty can expose CCAs to substantial financial risk by affecting the cost of market imports and, particularly, the value of export opportunities. This paper develops a robust optimization framework for CCA resource procurement that hedges against uncertainty in wholesale import and export prices. We extend the open-source GenX capacity expansion model with a budgeted robust optimization formulation that protects procurement decisions against adverse price realizations while remaining computationally tractable for large-scale, multistage planning studies. Using data from a California CCA, we show that market price uncertainty can materially change optimal procurement strategies. Rather than aggressively expanding renewable capacity based on expected export revenues, robust portfolios reduce dependence on uncertain market revenues, leading to more conservative renewable investments and less reliance on market exports. We demonstrate that moderate uncertainty budgets provide most of the risk-reduction benefits of robust planning at limited additional cost, whereas higher robustness levels yield diminishing benefits as portfolio conservatism increases. These findings provide guidance for selecting uncertainty budgets in IRP studies and demonstrate how robust optimization can support affordable and resilient procurement decisions as CCAs contribute to the transition toward low-carbon electricity systems.