Existing evaluations of AI models on high-finance work focus on retrieving answers from financial documents and producing numerically correct outputs. Real private equity (PE) and investment banking (IB) work also depends on choosing between conflicting sources, making sound assumptions, and keeping every step auditable. We introduce DeepFinance, an ultra long-horizon benchmark of 21 end-to-end PE and IB tasks that take a human professional an average of 27.1 hours. Each task pairs a request with a data room averaging 19 files and is graded against an average of 84 expert-written rubric criteria covering output, workflow, and judgment. We evaluate 22 model configurations, each in its native agent harness. The best configuration, GPT-6 Astra with max reasoning, satisfies 57.6% of rubric criteria, and the top four configurations are not statistically distinguishable. Frontier models often calculate correctly from the wrong inputs and find decision-relevant evidence without acting on it.
Binary decision-making under uncertainty arises across a wide range of applications, including energy systems scheduling, transportation, and workforce planning. While machine learning has been increasingly applied to accelerate or replace mixed-integer solvers in such settings, existing approaches rely on predictive learning, training models to reproduce optimal decisions from labeled data, which lacks knowledge of counterfactual outcomes and provides no guarantees that the resulting decisions will be feasible. This limitation is further compounded under distributional shift, where policies trained on historical data may produce systematically infeasible decisions. We propose a prescriptive learning framework for binary decision-making that addresses both limitations. Rather than predicting optimal labels, the framework jointly optimizes over linear decision policies and feasible assignments across all training scenarios, allowing feasible alternatives and their operational performance to be considered directly during training. We further extend the framework to a distributionally robust chance-constrained formulation using a Wasserstein ambiguity set, which enforces feasibility with a prescribed probability under the worst-case distribution in the ambiguity set and enables a tunable trade-off between robustness and solution quality. We demonstrate the framework on two case studies, namely a unit commitment problem and a staff allocation problem. Out-of-sample experiments under test distributions that either match or differ from the training distribution show that the robust model is more conservative when the two distributions coincide, while under distributional shift it substantially reduces unmet demand and lowers total cost by up to 76.5% relative to classical predictive learning.
In Bangla, a main verb is often followed by one of a small set of vector verbs. These verbs lose their own meaning and add a shade of meaning instead, such as completion. For example, likhe phello means that someone finished writing, not that they wrote and threw something. The same two verbs can also describe two separate events, so only the context shows which meaning is meant. This makes compound verbs a hard test of how language models handle multiword expressions. We present BanglaCV-Probe, a benchmark of 17 vector verbs with 1,044 minimal pairs, 204 entailment items and 197 pairs of correct and incorrect combinations, all checked by native speakers. We test 26 models (encoders, LLMs and translation systems) on four skills: telling the two meanings apart internally, understanding what a vector adds, preferring correct combinations and keeping the vector's meaning in translation. Probes on five encoders tell the two meanings apart and still reach up to 93% accuracy when trained on pairs without surface cues. Prompted models rarely use this knowledge. Entailment accuracy stays between 59% and 81%. Several LLMs answer Yes or No to almost every item, and nine of ten models miss that someone who ate up the rice left none. Most instruction-tuned LLMs translate vectors literally in 15-20% of vector sentences and in up to 55% for mora (die). On correct combinations, a small Bangla-only model (88%) matches BLOOM-7B1 (87%) and both beat Qwen2.5-7B (74%). Data and code: https://github.com/ashuvo25/Pole_OR_Vector
Fixed prompt-based language models face challenges in open-domain conversations, where versatile topics require different expert advice or demonstrations to maximize their universal performance. Motivated by the dynamic prompt studies in language generation, we extract different chatting strategies from varied forms of text corpus, which steer the language model to different styles based on human principles. We design and deploy a conversational companion called SCorPion, which uses GraphRAG to extract and summarize the knowledge graph globally, yield community-summarized strategies, and produce query-focused optimal strategies. SCorPion has strategic thinking, yields personified and stylized responses, and can be controlled by explicit factors. Code and further results can be found in https://bachozean.github.io/scorpion-page/.
Detectors of AI-generated text are usually tuned and tested on passages of several hundred tokens, yet many real cases are short. We study score-based detection as a function of passage length. We give a closed form for how the false-positive rate of a threshold set on long text changes as length falls, and a formula for the number of tokens a detector needs to reach a target true-positive rate. We calibrate thresholds within length bins with conformal p-values, and we add a bin-level rule that abstains when the detector cannot be informative at that length. On synthetic scores the formulas agree with simulation, and binned calibration holds the false-positive rate at its nominal value at every length. On three zero-shot detectors (log-likelihood, log-rank, Binoculars), three corpora and two open-weight generators, binned calibration gives a mean false-positive rate of 5.0% in every detector-length cell from 10 to 200 tokens, averaged over 300 random document-level resplits. A single threshold set at 200 tokens gives 29% for Binoculars at 10 tokens and about 1% for log-likelihood, as the closed form predicts. No detector reaches an 80% true-positive rate, and the best, Binoculars at 200 tokens, reaches 59%, so the rule abstains in every cell at that floor and releases only the 200-token bins at floors of 0.3 and 0.5.
Real-world deployment of federated learning (FL) requires resilience against adversarial clients and data privacy. Byzantine fault captures the worst-case behavior of adversarial clients, while differential privacy (DP) provides statistical guarantees against privacy loss. In particular, user-level DP guards a client's entire contribution and provides more realistic protection against information leakage. Despite intensive efforts, the minimax-optimal convergence bound of Byzantine-resilient FL with user-level DP remains open. We establish an information-theoretic lower bound for any Byzantine-resilient empirical risk minimization problem under user-level DP and heterogeneous clients. We propose ByzULDP, a central DP algorithm featuring a two-level momentum across the client and server. Client-side momentum mitigates stochastic gradient variance during robust aggregation. Unlike the fault-free case, naively combining non-linear Byzantine-resilient aggregation with central DP noise can be insufficient to guarantee convergence to a bounded neighborhood of a stationary point. Our key contribution is a server-side momentum that leverages a property specific to user-level DP, where privacy sensitivity depends on how client updates are constructed, to empower convergence. ByzULDP achieves the order-optimal convergence bound for strongly-convex objectives, with vanishing error terms decaying at the optimal O(1/T) and O(1/{T}) rates for strongly-convex and non-convex objectives, respectively. We corroborate our analysis with numerical experiments on real-world datasets across diverse privacy budgets.
Standard ensemble methods can improve robustness by mitigating individual model stochasticity, but they frequently fail to generalize beyond the training distribution, leaving systems vulnerable to out-of-distribution (OOD) data. We introduce the Robust Satisficing Ensemble (RSE), a scalable probabilistic aggregation framework designed to secure model combinations by minimizing their fragility against distributional changes. By bridging a robust satisficing objective with a quadratic risk surrogate, RSE derives an analytical inner adversary. This fundamental reformulation enables highly efficient weight updates without the computational bottleneck of standard minimax optimization, crucially decoupling optimization complexity from sample size to scale effectively to large datasets. In particular, we prove a geometrically decaying truncation error for the fragility approximation via Lagrange-Burmann expansion and establish stationarity guarantees for RSE. We evaluate RSE across a comprehensive spectrum of distribution shifts, including adversarial label shift (SST-5, TREC), natural subpopulation and temporal shifts (CivilComments, HuffPost). Our results demonstrate that RSE consistently outperforms strong robust baselines, including GroupDRO, IRM, and SRM. These findings establish RSE as a practical, theoretically grounded safeguard for reliable model aggregation in dynamic environments.
Pre-training large parametric "Vision-Language-Action" (VLA) models on large demonstration datasets has recently yielded promising results for general-purpose manipulation. However, these models still struggle under even small domain shifts relative to the training data, such as changes in camera viewpoint, background, object identity, or robot morphology. Motivated by the success of Retrieval-Augmented Generation (RAG) in language models, we pursue retrieval to boost VLA performance. Specifically, we seek to leverage retrieval from domain-mismatched, "in-the-wild" demonstration datasets like DROID to guide a VLA's behavior in a new target domain. To achieve this, we introduce RAGu, which overcomes robotics-specific challenges for RAG: (1) It incorporates simple geometric calibration to transform source demonstrations into the target setup and reduce cross-domain differences. (2) It uses a coarse-to-fine retrieval strategy that selects, from a large pool of pre-warped candidates, the trajectory most compatible with the base VLA. (3) It guides the VLA's flow-matching inference procedure rather than seeking to retrain the base model or expand its limited context window. We evaluate RAGu across real-world, simulation, sim-to-real transfer, and cross-embodiment humanoid manipulation. We find it improves real-world VLA success rates from 20% to 80% and outperforms the closest-related baselines in sim. Project website: raguvla.github.io
Human-AI complementarity captures the idea that a human assisted by an AI system can outperform both the human and the AI acting alone. In prediction tasks, this improvement is assessed against ground truth labels produced through a reference construction procedure and intended to represent the state of affairs to be predicted. Many human-AI interactions, including those where medical AI provides treatment recommendations, however, concern tasks other than predicting what obtains. Their relevant output concerns adequacy relative to a practical goal and depends on several evaluative dimensions. In this work, we examine the ground truth dependence of complementarity and extend it to human-AI interactions where ground truth is not the appropriate reference. We show that tasks beyond prediction can still admit a reference construction procedure that generates adequacy labels. We develop a route for extending complementarity to such settings. It replaces ground truth labels with context-dependent adequacy labels generated through a two-step scoring methodology inspired by multi-criteria decision analysis frameworks in healthcare and it makes the normative choices underlying reference construction explicit. Because the construction of adequacy labels embeds evaluative choices, we draw methodological inspiration from computational reliabilism to identify technical, scientific, and social factors relevant to supporting their construction. Our approach expands the empirical study of human-AI complementarity beyond prediction tasks, enabling its quantitative evaluation in decision-support settings where ground truth is not the appropriate evaluative reference.
Large Language Models (LLMs) exhibit approximate recall: they retain coarse associations about past events while losing fidelity on specific details. Past research on LLMs has quantified different models' forecasting ability in real-world events using prediction markets, but has not studied how losing said fidelity impacts their forecasting abilities. For example: while an LLM correctly recalls that the answer to: ''Did Anna Sawai win the Best Drama Actress in the 76th Emmy Awards?'' is ''Yes,'' it incorrectly states the event date as September 22 rather than September 15. In this paper, we introduce a novel metric called the Recall Fidelity Score to study how imprecision in internalized knowledge propagates to forecasting capability on future events. On resolved historical markets, we find that higher local recall fidelity around semantically similar past events predicts lower forecasting error on resolved markets, and the same signal predicts smaller distance from open markets prices.
Much of the automation handed to language-model agents is procedural: each step needs little intelligence, but a long unattended run must follow a skill or SOP stated once at the start. We show that agents drift from such procedures over long runs, and that restating the instructions or resuming a stopped agent only moves the drift to another shortcut. We study an executable finite-state machine (FSM) that keeps the workflow state outside the model's context and enforces it at three points: tool calls, state transitions and the end of the agent's turn. In a pre-registered experiment on 200-account SOP-Bench batches, Claude Haiku 4.5 with the FSM decided on average 139 accounts correctly and by the procedure, against 48 for the strongest baseline (95% CI of the difference [46, 130]), and kept this rate to the end; across three models, the gain tracked how much the plain agent drifted. Because the runtime, not the model, supplies this long-horizon reliability, the FSM can let small models carry procedural work: Haiku 4.5 with it averaged more such decisions than plain Claude Sonnet 4.6 at 38% of the cost. Analyses on SWE-bench Verified, τ^2-bench, and staged and cross-context workflows further suggest that reliable long-horizon automation can come from stronger execution control, not only stronger models.
Assessing climate risks requires projections of daily temperature distributions, including rare and spatially widespread extremes. Climate simulations are costly and biased, while correcting them toward observations risks suppressing inter-model differences. We train a single machine-learning emulator, a conditional diffusion model, on seventeen CMIP6 models and ERA5 reanalysis to generate spatially coherent daily temperature fields conditioned on model identity, date, and CO2-equivalent concentration. At sampling time, diffusion guidance reduces model biases by steering each emulated climate-model distribution toward observations, with its strength selected through perfect-model experiments using withheld simulated futures. Observational influence decreases with projection horizon and forcing. Across seventeen synthetic targets, guidance more than halves median biases in the 1st and 99th percentiles of daily temperature, improves probabilistic scores and reliability, and reduces median annual global-mean temperature RMSE by 32-57% across three emission scenarios. In a detailed benchmark, guidance also outperforms established bias corrections on daily-field probabilistic scores. Efficient sampling generates large ensembles without further climate simulations. For example, our framework corrects CMIP6's roughly threefold historical overestimate of extreme-heat coverage, yet projects that concurrent extremes will affect nearly half of global land on a typical day in the 2080s under SSP3-7.0.
Continual model merging incrementally integrates task-specific models without revisiting past models or data. However, repeated merging entangles past task components and distorts their singular geometry, making balanced task-wise allocation increasingly difficult. We propose Global Singular Subspace Separation and Restoration (GS3R), which augments the merged update with a compact restoration state. At each merge, GS3R restores the retained singular geometry, separates task-wise subspaces, and globally reallocates the available rank across tasks fairly. To reduce merging-time memory overhead, we further introduce Pre-merge Restoration and Flushing (PRF). Experiments on vision, NLP, and multimodal benchmarks show that GS3R provides a strong accuracy-efficiency trade-off.
Sampling-based model predictive control (MPC) is treated as a family of distinct algorithms: predictive sampling, MPPI, the cross-entropy method, and their biased, colored-noise, and annealed variants all sample control sequences and reweight them by cost. We show a single Sequential Monte Carlo (SMC) inference algorithm implements them all, specified by three independent axes: the proposal, the update rule that reduces cost-weighted particles to a nominal plan, and the effective-sample-size threshold for resampling. We evaluate SMC-based planning on MuJoCo non-linear control, open-loop navigation through obstacle fields, 7-DoF motion planning, and non-prehensile push-T, in simulation and on a Franka FR3. Resampling combats particle degeneracy at long horizons. At a 28-knot horizon in our navigation domain it cut the collision rate by 30% over non-resampling baselines. Programmable proposals improve push-T success over perturbative baselines, improving from 1% to 100% in simulation and 10% to 90% on hardware.
Reinforcement learning for large language model post-training often reuses each batch of sampled rollouts for several gradient passes to amortize the cost of generation. Under such multi-pass reuse, existing objectives fail in different ways: unclipped updates collapse as the policy drifts away from the sampling policy, adding a negative baseline avoids collapse but reinforces every sampled response, and PPO-style ratio clipping cuts off gradients abruptly. We first ask why reuse is fragile. Measuring how well an update estimated from a small rollout group aligns with the KL-regularized target on held-out rollouts of the same prompts, we find that this alignment drops by nearly an order of magnitude from in-sample to held-out rollouts across three base models, suggesting that repeated passes largely overfit the sampled group. Motivated by this finding, we propose Bounded Residual-Aligned Training (BRAT), motivated by a proximal update toward a bounded reweighting of the sampling policy. BRAT assigns each token a gradient weight that decreases linearly with its probability ratio and vanishes at a trust-region boundary, and normalizes group-centered rewards by their maximum magnitude, keeping advantages zero-mean and bounded without extra hyperparameters. With four reuse passes, BRAT avoids training collapse, limits policy drift, and achieves higher average accuracy with lower seed variance on most benchmarks.
We study the problem of controlling instability in the Vlasov--Poisson plasma systems, a fundamental challenge for nuclear fusion control. Motivated by the partial observability of plasma states in practice, we investigate imitation learning algorithms that learn from a fully-observable expert controller, but operate under macroscopic measurements. We theoretically establish a separation between online and offline imitation learning --- even a small behavior cloning error leads to error compounding that exponentially amplifies instability over time. By way of contrast, the online imitation learning loss can polynomially control the stability of rollout trajectories. We then propose a DAgger-style algorithm that learns the stabilizing controller under partial observability. Simulation results on a 1D Vlasov--Poisson system demonstrate that our algorithm can effectively stabilize plasma instabilities over longer time horizons than behavior cloning.
Traditional text-based person retrieval (TPR) assumes that a witness provides a complete description at once, whereas interactive person retrieval (Inter-PR) refines a partial description through question-answer exchanges. In this setting, the questioner inspects retrieved candidates to decide what evidence to request next, while the answerer sees only the witness-side memory and cannot report candidate-specific visual differences. Existing methods mainly learn offline question orderings, leaving this visual-feedback loop underexplored. To address this limitation, we introduce visual-feedback interactive person retrieval (VFI-PR), a new interactive retrieval paradigm in which retrieved candidates produce visual feedback that becomes evidence for subsequent questioning and re-ranking. To support training and evaluation in this setting, we construct VFI-PEDES, a dataset and evaluation protocol for visual-feedback interaction, and propose VFI-ReID, a framework that learns to use visual feedback during interactive person retrieval. Through the Dynamic Accumulated-evidence Retriever (DAR) and retrieval-aware question optimization, VFI-ReID adapts retrieval and question generation to evolving evidence states, turning visual feedback into complementary evidence for subsequent retrieval. Experiments show that VFI-ReID consistently improves retrieval across VFI-PR interaction stages, demonstrating the utility of visual feedback for interactive person retrieval and outperforming existing Inter-PR baselines.
