Research Papers

Pre-training followed by fine-tuning has become the dominant recipe for learning performant policies, and in value-based reinforcement learning (RL) this raises a natural question: given a pretrained policy, should the Q-function be pretrained on offline data too? Conventional wisdom suggests it should, but recent results show that online RL with a randomly-initialized Q-function can result in highly performant and reliable policies without needing to pretrain the Q-function. In this paper, we systematically study whether pretraining the Q-function actually helps when fine-tuning on top of a pretrained base policy. We find, surprisingly, that naive Q-function pretraining often provides little benefit over random initialization. We show this stems from a fundamental mismatch: the Q-function learned during pretraining targets the pretrained policy's Q-function, not the Q-function that online fine-tuning converges to, and this gap persists even after offline value maximization. Motivated by this finding, we propose Initialization via Policy Ensemble (IPE), a simple method that trains multiple diverse policies and uses their pooled rollouts to bootstrap the Q-function learning in online RL. Across a suite of challenging continuous control benchmarks, IPE yields an average 1.26x improvement in fine-tuning performance over naive Q-function pre-training.

We present a novel approach to regression tasks using classification which is motivated by the mechanism used by fruitflies to sense their environment. Specifically, we formulate a general framework for learning nonlinear input-output relationships by replacing complex global surrogate models with a finite library of representative local patterns. Since scientific data often occupy limited and recurring regions of the input space, we generate predictions by measuring similarities between a query and stored patterns, then combining their associated responses through weighted reconstruction. We apply this approach to nonlinear dynamical systems, data-driven regression, and physics-informed learning using suitable embeddings and similarity measures. For dynamical systems, our offline-online workflow extracts patterns from data or governing equations during the offline phase, while online prediction requires only similarity evaluation and response aggregation. This structure helps us reduce computational and memory demands while providing explicit control over the trade-off among accuracy, storage, and inference cost.

Forecasts of explosive AI progress hinge on AI agents automating AI research. But evidence on whether agents can carry out open-ended AI research is thin. Current evaluations either test agents on narrow, verifiable tasks, which excludes open-ended research, or submit AI-generated papers to blind peer review, which is overstretched, stochastic, and suffers from poor review quality. We introduce a third way to measure progress towards AI R\&D automation. An agent takes on the central, open-ended research question of a high-quality unpublished paper, and the paper's original authors grade its output. We call these shadow evaluations. We ran shadow evaluations on two unpublished NeurIPS 2026 submissions, giving frontier agents six days and thousands of dollars of compute. The agents completed all of the engineering without human help, yet could not make substantial progress towards answering the research questions. As a result, both papers were unambiguously rejected by the authors. We identify five recurring failure modes: poor judgment about the bar for publishable research, uncreative responses to shortcomings in the research design, ineffective backtracking from dead ends, poor resource awareness, and instruction drift. A robustness check with a second model and scaffold reproduced these failures. We release the expert reviews, survey responses, agent repositories, and logs. Our results provide early evidence that today's agents can do the engineering of AI research, but struggle with critical parts of the research lifecycle.

We introduce APEX-Accounting, a benchmark built by Mercor in partnership with Ramp, to assess whether frontier models can do the real work of accountants. Tasks include reconciling accounts, accruing expenses, posting transactions, and producing reports. The private eval set comprises 160 tasks, split across 10 worlds. Each world contains an accounting system, as well as spreadsheets, PDFs, and other files. Every task was authored and solved by experts in accounting and bookkeeping, who also wrote grading rubrics. Across nine frontier models, Claude-Fable-5 (Max) leads with 56.4% Mean Criteria@3, ahead of Muse-Spark-1.1 (xHigh) at 52.6%. No model scores more than 2.6% Pass^8 (GPT-5.6-Sol (Max+Pro)) and the highest Pass@8 is 21.5% (Muse-Spark-1.1 (xHigh)). We experiment with increasing the token budget from $1 to $50 and observe an instance of Simpson's paradox: scores increase as the token budget increases but within a given budget-constrained harness, scores are lower on tasks where the model spends more tokens. As APEX-Accounting is a closed benchmark, leaderboard evals can be run for any frontier model on request.

Accurate option prices do not imply accurate recovery of the latent risk-neutral density. We study this distinction with two complementary benchmarks. A controlled benchmark exposes simulator-truth densities for latent evaluation, while a chronological NIFTY benchmark tests only held-out market prices. A two-component lognormal mixture has the lowest aggregate price, $L^1$, Wasserstein, and fixed-tail errors on the synthetic benchmark. Learned operators retain narrower strengths: DeepONet reduces 1% quantile and variance error by 39.0% and 34.6% relative to the mixture, and a quote transformer reduces $L^1$ by 16.4% on the structurally misspecified Merton family. A numerical conditioning analysis explains why these rankings can differ: after enforcing mass and forward constraints, 95 of 126 pricing directions are numerically null, and two densities separated by $L^1 = 0.061$ produce identical prices on the covered strikes. On 524 held-out NIFTY calls, validation-selected test-time adaptation reduces DeepONet RMSE by 28.3%, but per-expiry mixture and SVI fits remain much more accurate. The evidence supports target-dependent inductive bias, not a universal winner.

Conversational AI is increasingly positioned as a teammate rather than a tool, yet we know little about how its presence reshapes communication among the humans on the team. We examined sociocognitive communication dynamics in team decision-making using Group Communication Analysis (GCA), team surveys, and lexical analyses of team discourse. Teams completed a high-stakes moral-dilemma decision task in a randomized controlled study: 16 teams of two students plus an AI teammate, and 17 all-human teams of three. Across six GCA dimensions and survey outcomes, we find that the AI teammate was the single most talkative and self-cohesive member of every treatment team, yet its contributions carried the least new information and the lowest density. The presence of AI also reshaped communication amongst humans. In AI-human teams, human teammates showed lower responsivity and social impact toward one another and reported lower levels of belonging and status. Greater AI dominance in the conversation was associated with students feeling less valued as team members. Additionally, this social cost is immediate and present at baseline; it does not emerge over the course of the conversation. Drawing on these results, we discuss a research agenda extending to voice-based and longitudinal settings.

Effective collaboration with novel and diverse partners is a crucial skill for autonomous agents. Most current ad-hoc teamwork (AHT) approaches assume that agents will collaborate on a single, fixed task and that the partner's capabilities, their ability to successfully execute the desired action, are already known. In reality, a partner's true capabilities are often hidden, and human collaborators may act sub-optimally on tasks with multiple valid strategies. To address these limitations, we extend ad-hoc teamwork into a multi-task setting by re-framing it as a problem of joint planning with decentralised execution under hidden partner capabilities. We introduce CE-CM (Capability Estimation via Contextual Models), an approximate Bayesian method that infers task-invariant capability vectors. By using simulation-based sampling, the agent estimates capabilities and induces a contextual Multi-agent Markov Decision Processes for planning. This approach requires no population pre-training and refines its beliefs online from just a few tasks. To account for human unpredictability, we propose CE-CM-Div, an extension that evaluates capability hypotheses against diverse planner rollouts rather than a single optimal trajectory. Simulated experiments demonstrate that CE-CM rapidly recovers hidden capabilities, reduces infeasible action assignments, and adapts to changes over time. Furthermore, in an offline human study of 225 trajectories from 15 participants, CE-CM-Div substantially improved capability estimates over the baseline CE-CM method. Our results suggest capability-based modelling is a promising interpretable, task-agnostic representation in the studied settings, demonstrating that accounting for behavioural diversity is essential for robust human-AI teaming.

Traditional search systems are optimized to retrieve items that strictly match a query, often prioritizing precision over recall. In e-commerce marketplaces and particularly grocery, this paradigm is limiting, as user satisfaction and commercial outcomes depend heavily on the discoverability of substitute, complementary, and thematically related items. In this paper, we present a scalable system for discovery-augmented search that leverages intent-conditioned recall expansion. Our approach generates implicit user intents to expand candidate recall while maintaining relevance. The system addresses the cost-quality tradeoff of generative retrieval through a two-stage hybrid architecture. First, we leverage closed-weight large language models (LLMs) to maximize discoverability for head queries. To extend these benefits to tail queries, we then introduce a finetuned small language model (SLM), trained via LoRA adapters and teacher-student distillation. We evaluate the system using a rigorous dual framework: (a) LLM-as-a-judge metrics validated against human preferences for semantic quality, and (b) end-to-end session-level purchase analysis. Results demonstrate that our approach improves both intent generation quality and downstream retrieval effectiveness, extending discovery coverage from approximately 60% to 80% of query traffic at roughly 30% of the teacher model's inference cost, offering a viable path for deployment in large-scale marketplaces. Beyond relevance gains, discovery-augmented search may serve as a marketplace-balancing mechanism, giving long-tail and emerging supply an opportunity for query-conditioned exposure.

Solving a continuous algebraic constraint system requires two decisions: which values satisfy the constraints, and which structural augmentation renders an unsolvable system solvable. Classical solvers answer the first well and the second only by enumeration. On that discrete decision, a candidate-conditioned repair ranker choosing among K augmentations reaches the exhaustive-search ceiling at a fraction of the calls, outperforming random (0.997 vs 0.236 balanced nonlinear menu accuracy; p < 10^-70; 0.982 +/- 0.006 across seeds) and beating a budget-matched per-candidate probe on accuracy and cost. MARC turns such a system into a factor graph, over which a graph-neural diffusion denoiser proposes assignments, descent on an exact computer-algebra energy polishes them, and an exact symbolic checker certifies solutions. Evaluations of diffusion-based proposals rarely include one control: random multi-start under the same refinement budget. Applied to our system, it sharply curtails what the learned proposal contributes on the value decision. Does it beat random multi-start at choosing satisfying assignments? Only narrowly, in a predictable regime. Across trapped low-dimensional families it ties with random restart, but dominates in high dimension, where random search fails. Once variables couple, the advantage is gone. Since all methods share one polish and one checker, best-of-K random multi-start succeeds with probability exactly 1 - (1 - q(n))^K, where q(n) is single-start reachability; one measured constant, with no free parameters, reproduces the entire curve (mean absolute error 0.012). The favorable regime is not specific to our synthetic families: across eight real-world systems in robotics, positioning, optimization, and algebra, classical multi-start solved all eight, none in the learning-favorable regime. We map the regimes in which learned proposals improve solvers.

Large language model (LLM) agents are increasingly expected to assist users in completing tasks. However, existing benchmarks provide limited support for evaluating whether agents can carry out office-suite workflows at a reasonable cost. We introduce OmegaUse-OfficeVal, a benchmark for evaluating LLM agents on long-horizon office-suite tasks with task-level economic grounding. The benchmark comprises 100 tasks derived from office-suite requests proposed by practitioners and adapted through a privacy-preserving process. On average, these tasks require 2.32 hours of human labor to complete. An important feature of the benchmark is that each task is paired with two economic signals: human labor time and task price proxy. These signals enable direct comparisons between human costs and LLM inference costs, as well as value-weighted evaluation. To support stable evaluation, we develop code-based verifiers from fine-grained rubrics. We evaluate several frontier LLMs together with a human baseline. Although all evaluated LLMs are substantially cheaper and faster than human workers, they have not yet approached human-level deliverable quality. The code and dataset are fully open-sourced, and more information is available on our project website: https://omegause-officeval.github.io.

CT vision-language foundation models have demonstrated promising performance across downstream tasks, but are typically trained with whole-volume representations that dilute fine-grained anatomical signals. Fine-grained vision-language pre-training addresses this by aligning anatomy-level visual features with anatomy-specific text, but in doing so discards the global context that whole-volume models provide. Furthermore, existing fine-grained approaches train from scratch, making them computationally expensive. We introduce Anatomy Contextualized Adaptation (ACA), a lightweight framework that adapts frozen CT foundation model representations for anatomy-level vision-language alignment while enhancing global contextualization. ACA uses TotalSegmentator to decompose CT volumes into anatomy-level embeddings, which are refined via a transformer that captures cross-anatomy relationships, and aligned to both per-anatomy and scan-level text extracted from radiology reports. Evaluated on Merlin and CT-RATE, ACA consistently outperforms both the frozen foundation model baselines and existing fine-grained methods in zero-shot finding classification, while requiring less than one hour of training once embeddings are cached. The attention weights learned by ACA's inter-anatomy transformer additionally indicate plausible cross-anatomy context routing. Altogether, these results support ACA as a lightweight approach for adapting CT foundation models to anatomically grounded vision-language alignment while preserving and enhancing global anatomical context.

Accurate intraday forecasts of offshore wind are becoming increasingly important for power system operation and the integration of growing shares of offshore wind energy. Operational forecasts rely predominantly on numerical weather prediction (NWP), which is not optimized for lead times of minutes to hours, where initial-condition accuracy dominates forecast skill. Although satellite scatterometer observations are routinely assimilated into NWP, they have not previously been used directly for forecasting. Here we present WindCastNet, the first satellite-based nowcasting framework for offshore wind speed and direction, introducing a new paradigm for intraday forecasting that learns from spatiotemporally irregular satellite observations. WindCastNet predicts offshore wind fields from observations acquired by satellite scatterometer constellations. WindCastNet employs a partial convolutional long short-term memory network that exploits microwave radar observations from the European, Chinese, and Indian scatterometers despite their irregular spatial coverage, asynchronous sampling, and variable revisit times. Spatial observation masks and inter-observation intervals are encoded, while a continuous temporal representation enables forecasts at arbitrary lead times. Evaluated over the North Sea, WindCastNet reduces the root-mean-square error by 23% and 7% relative to the HARMONIE MEPS model at lead times of 1 and 2 h, respectively, and outperforms persistence by 9-15% during the first three forecast hours. Forecast skill decreases under strong-wind conditions and spatially non-uniform flow. These results demonstrate that satellite scatterometer constellations can provide an independent and competitive source of short-term offshore wind forecasts, opening new opportunities for renewable energy forecasting but also broader marine weather applications, including tropical cyclone nowcasting.

Coding agents have made substantial progress on software engineering tasks that modify existing codebases, including bug fixing and feature implementation. However, constructing a complete program from scratch remains a major challenge: even the frontier models evaluated on ProgramBench fully resolve fewer than 1% of tasks. One obstacle is the lack of scalable training environments for this from-scratch setting, spanning the whole software engineering life cycle, as existing environment-construction frameworks focus only on a single phase in software development. To address this gap, we introduce MindForge, an automated pipeline that converts open-source command-line programs into source-free environments that expose only a compiled reference executable and its documentation. Using MindForge, we construct training environments from repositories disjoint from those in ProgramBench, and curate a high-quality data recipe consisting of program synthesis trajectories using GLM-5.2 as the teacher agent. Fine-tuning Qwen3.6-27B on these trajectories increases its ProgramBench average test pass rate from 37.98% to 49.51%, achieving performance comparable to substantially larger frontier models. Moreover, the fine-tuned model consistently improves over the base model across all seven unseen software engineering benchmarks, spanning long-horizon repository generation and translation, bug fixing, feature implementation, and cross-language issue resolution, with absolute gains of 31.00 points on RepoZero-C2Rust, 14.16 on DeepSWE, 10.70/4.56 on NL2Repo-Bench (with/without tests), 5.04 on SWE-bench Verified, 5.93 on SWE-bench Pro, 5.22 on SWE-bench Multilingual, and 4.94 on FeatBench.

High-stakes decision systems in credit scoring, fraud detection, healthcare, and industrial safety require reliable uncertainty quantification under severe class imbalance and asymmetric error costs. Standard marginal conformal prediction (CP) provides valid overall coverage guarantees; however, we show that it severely under-covers rare, costly minority classes, with minority-class coverage dropping to as low as 0.5% on certain datasets. To characterize and address this limitation, we conduct a comprehensive benchmark comparing marginal CP, class-conditional (Mondrian) CP, and cost-controlled abstention mechanisms across 15 real-world imbalanced tabular datasets, 7 classification models, 3 probability calibration techniques, and 10 random seeds, resulting in 3,150 experimental runs. Our results show that Mondrian CP restores valid minority-class coverage, achieving an average minority-coverage improvement of 61.7 percentage points over marginal CP (p < 1e-80). Furthermore, combining Mondrian CP with cost-controlled abstention significantly reduces expected decision cost compared with standard decision boundaries, confidence-based rejectors, and risk-controlled rejectors under realistic human review budgets. We further quantify dataset-specific break-even thresholds at which deferring ambiguous instances to human experts becomes cost-effective. These findings provide practical guidance for deploying distribution-free, cost-aware uncertainty quantification in high-stakes decision support systems.

This project aimed to develop a novel reservoir compute (RC) implementation framework targeting high-speed operation and integration with CMOS digital logic. With the target workload of branch prediction (BP) for multistage pipelined central pro-cessing unit (CPU) cores. For this, a novel memristor based RC design framework was developed within the context of the workload requirements. This was then implemented in simulation using industry standard modelling languages of System Verilog (SV) and Verilog-AMS (VAMS).The developed RC design framework was subsequently verified using a basic sequence detection task before further benchmarking for its effectiveness at BP. The developed RC framework was tested using the Dhrystone performance benchmark, while targeting the RISC-V RV64GC instruction set architecture (ISA). Conducted testing demonstrates that RC shows great promise for ap-plication to BP and is capable of achieving impressive overall prediction accuracy. However, testing also shows that further refinement of the developed RC design framework is necessary to address shortfalls in the adaptability of the proposed RC system. As comparison against the state of the art TAGE predictor showed the proposed RC design framework to be 15x slower to adapt to changes in branching behaviour.

Vision-language-action (VLA) models remain constrained by scarce action-labeled robot data, whereas action-free videos offer abundant observations of physical change. Latent action models can extract such priors, but reconstruction-trained codes may predict future observations without the structure required for joint generation with robot actions. Existing structured methods add temporal constraints but retain deterministic transition points, so residual errors in locally inferred transitions may propagate and compound under recursive composition. We introduce DLAM, a distributional latent-action model that represents each transition as a diagonal Gaussian. Reconstruction conditioned on the reference frame grounds the mean in observed visual change, while normalized composition and reversal over equal-gap triplets constrain both the mean and dimension-wise variance. Variance composition uses a lightweight shared-correlation coefficient to account for dependence between adjacent transitions that share an intermediate frame, whereas reversal negates the mean and preserves the variance. For downstream policy learning, we freeze the encoder and train a flow-matching policy to jointly generate mean transition sequences and robot actions. On held-out transitions, DLAM learns more temporally consistent latent dynamics than existing latent-action baselines and achieves stronger direct and cumulative reconstruction on held-out videos. Under the same controlled $π_0$ transfer protocol, it also improves policy performance on MetaWorld MT50, LIBERO, and real-world manipulation tasks. Controlled ablations show that normalized mean constraints account for most of the reconstruction gain, while learned variance and correlation-aware composition provide complementary improvements in downstream control.

Writing and communication are increasingly mediated by large language models (LLMs) that are being used to draft, revise and polish text. Although such assistance can improve clarity and help authors meet institutional expectations, widespread reliance on shared models may reduce population-level variation in linguistic form, a phenomenon we refer to as linguistic monoculture. We develop a mathematical framework in which authors and LLMs are represented as distributions over linguistic features and coevolve through repeated interaction. We analyze three interaction mechanisms: a shared model with a fixed linguistic distribution, a shared model recursively updated from author outputs, and personalized models updated through author-specific and population-level feedback. We characterize the resulting equilibria and convergence rates, showing that, shared models can drive authors toward a common norm, recursive feedback relocates the shared norm without altering pairwise spread under common conformity, and personalization can preserve a family of distinct author-model equilibria with nonzero linguistic diversity. We then endogenize conformity as a strategic choice trading off private benefits from clarity, legibility, and perceived fluency against distinctive style. Within this utility model, individually rational authors may conform more than is socially optimal because they do not internalize the value their distinctiveness provides to others, creating a negative externality and a price of monoculture that is finite for each fixed instance but can grow without bound when distinctiveness dominates authenticity. Synthetic simulations illustrate how fixed shared assistance, recursive feedback, and personalization produce different long-run diversity outcomes.

An agent acting under partial observability must retain a recursively updateable statistic of history that restores the Markov property, but the smallest such statistic is generally unknown. We characterize this minimal Markov sufficient statistic for holonomy-cover decision processes, a structured POMDP class in which the visible dynamics are Markov and every realized visible transition applies a fixed permutation to a hidden mode. In particular, we construct the stable quotient, the coarsest observation-wise abstraction preserving one-step rewards and quotient successors, and prove that the pair of the current observation and stable class forms an exact finite Markov state. When the current class is correctly initialized, exact class tracking requires exactly the minimal memory symbols, in the sense that under reachability and pairwise decision separation at a maximizing observation, no arbitrary finite-memory controller can use fewer. Under resettable diagnostics, nearest-prototype class inference has exponentially decaying error, and a calibrate-then-restart reduction transfers finite-MDP guarantees to the recovered state. The results enable \emph{Holonomy Memory Reinforcement Learning}. It represents memory by the current stable class, updates it through ordered edge transports, identifies local class coordinates when diagnostics are available, and applies a standard finite-MDP RL backbone after synchronization. Experiments recover an exact compression from raw states to quotient states and achieve perfect paired-order accuracy with three decision-time memory states, matching the quotient oracle and outperforming the non-oracle baselines.