arXiv subjects

Osbert Bastani

Publications and source records attributed to Osbert Bastani.

At least 37 records · Page 2Linked to original sources

arXiv2026

SPARLING: Learning Latent Representations with Extremely Sparse Activations

Real-world processes often contain intermediate state that can be modeled as an extremely sparse activation tensor. In this work, we analyze the identifiability of such sparse and local latent intermediate variables, which we call motifs. We prove our Motif Identifiability Theorem, stating that under certain assumptions it is possible to precisely identify these motifs exclusively by reducing end-to-end error. Notably, we do not assume identifiability of parameters, but rather of a latent intermediate representation output by a local model, thus allowing these representations to be arbitrarily complex functions of the input. Additionally, we provide the Sparling algorithm, which uses a new kind of informational bottleneck that enforces levels of activation sparsity unachievable using other techniques. We confirm empirically that extreme sparsity is necessary to achieve good intermediate state modeling. On synthetic domains, we are able to precisely localize the intermediate states up to feature permutation with > 90% accuracy, even though we only train end-to-end.

cs.LG↗

arXiv2026

Adversarial Query Synthesis via Bayesian Optimization

Benchmark workloads are extremely important to the database management research community, especially as more machine learning components are integrated into database systems. Here, we propose a Bayesian optimization technique to automatically search for difficult benchmark queries, significantly reducing the amount of manual effort usually required. In preliminary experiments, we show that our approach can generate queries with more than double the optimization headroom compared to existing benchmarks.

cs.DB↗

arXiv2026

Evaluating the Diversity and Quality of LLM Generated Content

Recent work suggests that preference-tuning techniques -- such as Reinforcement Learning from Human Feedback (RLHF) methods like PPO and GRPO, as well as alternatives like DPO -- reduce diversity, creating a dilemma given that these models are widely deployed in applications requiring varied outputs. We argue that diversity without consideration of quality has limited practical value. To address this issue, we introduce a framework for measuring effective semantic diversity -- diversity among outputs that meet quality thresholds -- which better reflects the practical utility of large language models (LLMs). Using open-ended tasks that require no human intervention, we find counterintuitive results: when using diversity metrics that do not explicitly consider quality, preference-tuned models -- particularly those trained via RL -- often produce outputs with lower diversity; however, these same preference-tuned models generate greater effective semantic diversity than supervised fine-tuned (SFT) or base models. Our analysis further shows another trend: while larger models may exhibit greater effective semantic diversity than smaller models, the smaller models are consistently more parameter-efficient at producing unique content within a fixed sampling budget. These findings have practical implications for applications that require diverse yet high-quality outputs, from creative assistance to synthetic data generation.

cs.CL↗

arXiv2026

Prior-Agnostic Incentive-Compatible Exploration

In bandit settings, optimizing long-term regret metrics requires exploration, which corresponds to sometimes taking myopically sub-optimal actions. When a long-lived principal merely recommends actions to be executed by a sequence of different agents (as in an online recommendation platform) this provides an incentive misalignment: exploration is "worth it" for the principal but not for the agents. Prior work studies regret minimization under the constraint of Bayesian Incentive-Compatibility in a static stochastic setting with a fixed and common prior shared amongst the agents and the algorithm designer. We show that (weighted) swap regret bounds on their own suffice to cause agents to faithfully follow forecasts in an approximate Bayes Nash equilibrium, even in dynamic environments in which agents have conflicting prior beliefs and the mechanism designer has no knowledge of any agents beliefs. To obtain these bounds, it is necessary to assume that the agents have some degree of uncertainty not just about the rewards, but about their arrival time -- i.e. their relative position in the sequence of agents served by the algorithm. We instantiate our abstract bounds with concrete algorithms for guaranteeing adaptive and weighted regret in bandit settings.

cs.GT↗

arXiv2026

Synthesizing Trajectory Queries from Examples

Data scientists often need to write programs to process predictions of machine learning models, such as object detections and trajectories in video data. However, writing such queries can be challenging due to the fuzzy nature of real-world data; in particular, they often include real-valued parameters that must be tuned by hand. We propose a novel framework called Quivr that synthesizes trajectory queries matching a given set of examples. To efficiently synthesize parameters, we introduce a novel technique for pruning the parameter space and a novel quantitative semantics that makes this more efficient. We evaluate Quivr on a benchmark of 17 tasks, including several from prior work, and show both that it can synthesize accurate queries for each task and that our optimizations substantially reduce synthesis time.

cs.PL↗

arXiv2026

Optimal Program Synthesis via Abstract Interpretation

We consider the problem of synthesizing programs with numerical constants that optimize a quantitative objective, such as accuracy, over a set of input-output examples. We propose a general framework for optimal synthesis of such programs in a given domain specific language (DSL), with provable optimality guarantees. Our framework enumerates programs in a general search graph, where nodes represent subsets of concrete programs. To improve scalability, it uses A* search in conjunction with a search heuristic based on abstract interpretation; intuitively, this heuristic establishes upper bounds on the value of subtrees in the search graph, enabling the synthesizer to identify and prune subtrees that are provably suboptimal. In addition, we propose a natural strategy for constructing abstract transformers for monotonic semantics, which is a common property for components in DSLs for data classification. Finally, we implement our approach in the context of two such existing DSLs, demonstrating that our algorithm is more scalable than existing optimal synthesizers.

cs.PL↗

arXiv2026

Winner's Curse Drives False Promises in Data-Driven Decisions: A Case Study in Refugee Matching

A major challenge in data-driven decision-making is accurate policy evaluation-i.e., guaranteeing that a learned decision-making policy achieves the promised benefits. A popular strategy is model-based policy evaluation, which estimates a model from data to infer counterfactual outcomes. This strategy is known to produce unwarrantedly optimistic estimates of the true benefit due to the winner's curse. We searched the recent literature on data-driven decision-making, identifying a sample of 55 papers published in the Management Science in the past decade; all but two relied on this flawed methodology. Several common justifications are provided: (1) the estimated models are accurate, stable, and well-calibrated, (2) the historical data uses random treatment assignment, (3) the model family is well-specified, and (4) the evaluation methodology uses sample splitting. Unfortunately, we show that no combination of these justifications avoids the winner's curse. First, we provide a theoretical analysis demonstrating that the winner's curse can cause large, spurious reported benefits even when all these justifications hold. Second, we perform a simulation study based on the recent and consequential data-driven refugee matching problem. We construct a synthetic refugee matching environment (calibrated to closely match the real setting) but designed so that no assignment policy can improve expected employment compared to random assignment. Model-based methods report large, stable gains of around 60% even when the true effect is zero; these gains are on par with improvements of 22-75% reported in the literature. Our results provide strong evidence against model-based evaluation.

stat.ML↗

arXiv2026

Beating the Winner's Curse via Inference-Aware Policy Optimization

There has been a surge of recent interest in automatically learning policies to target treatment decisions based on rich individual covariates. In addition, practitioners want confidence that the learned policy has better performance than the incumbent policy according to downstream policy evaluation. However, due to the winner's curse -- an issue where the policy optimization procedure exploits prediction errors rather than finding actual improvements -- predicted performance improvements are often not substantiated by downstream policy evaluation. To address this challenge, we propose a novel strategy called inference-aware policy optimization, which modifies policy optimization to account for how the policy will be evaluated downstream. Specifically, it optimizes not only for the estimated objective value, but also for the chances that the estimate of the policy's improvement passes a significance test during downstream policy evaluation. We mathematically characterize the Pareto frontier of policies according to the tradeoff of these two goals. Based on our characterization, we design a policy optimization algorithm that estimates the Pareto frontier using machine learning models; then, the decision-maker can select the policy that optimizes their desired tradeoff, after which policy evaluation can be performed on the test set as usual. Finally, we perform simulations to illustrate the effectiveness of our methodology.

stat.ML↗

arXiv2026

Are AI Capabilities Increasing Exponentially? A Competing Hypothesis

Rapidly increasing AI capabilities have substantial real-world consequences, ranging from AI safety concerns to labor market consequences. The Model Evaluation & Threat Research (METR) report argues that AI capabilities have exhibited exponential growth since 2019. In this note, we argue that the data does not support exponential growth, even in shorter-term horizons. Whereas the METR study claims that fitting sigmoid/logistic curves results in inflection points far in the future, we fit a sigmoid curve to their current data and find that the inflection point has already passed. In addition, we propose a more complex model that decomposes AI capabilities into base and reasoning capabilities, exhibiting individual rates of improvement. We prove that this model supports our hypothesis that AI capabilities will exhibit an inflection point in the near future. Our goal is not to establish a rigorous forecast of our own, but to highlight the fragility of existing forecasts of exponential growth.

cs.AI↗

arXiv2026

Knowledgeable Language Models as Black-Box Optimizers for Personalized Medicine

The goal of personalized medicine is to discover a treatment regimen that optimizes a patient's clinical outcome based on their personal genetic and environmental factors. However, candidate treatments cannot be arbitrarily administered to the patient to assess their efficacy; we often instead have access to an in silico surrogate model that approximates the true fitness of a proposed treatment. Unfortunately, such surrogate models have been shown to fail to generalize to previously unseen patient-treatment combinations. We hypothesize that domain-specific prior knowledge - such as medical textbooks and biomedical knowledge graphs - can provide a meaningful alternative signal of the fitness of proposed treatments. To this end, we introduce LLM-based Entropy-guided Optimization with kNowledgeable priors (LEON), a mathematically principled approach to leverage large language models (LLMs) as black-box optimizers without any task-specific fine-tuning, taking advantage of their ability to contextualize unstructured domain knowledge to propose personalized treatment plans in natural language. In practice, we implement LEON via 'optimization by prompting,' which uses LLMs as stochastic engines for proposing treatment designs. Experiments on real-world optimization tasks show LEON outperforms both traditional and LLM-based methods in proposing individualized treatments for patients.

cs.LG↗

arXiv2026

Opportunistically Parallel Lambda Calculus

Scripting languages are widely used to compose external calls such as native libraries and network services. In such scripts, execution time is often dominated by waiting for these external calls, rendering traditional single-language optimizations ineffective. To address this, we propose a novel opportunistic evaluation strategy for scripting languages based on a core lambda calculus that automatically dispatches independent external calls in parallel and streams their results. We prove that our approach is confluent, ensuring that it preserves the programmer's original intent, and that it eventually executes every external call. We implement this approach in a scripting language called Opal. We demonstrate the versatility and performance of Opal, focusing on programs that invoke heavy external computation through the use of large language models (LLMs) and other APIs. Across five scripts, we compare to several state-of-the-art baselines and show that opportunistic evaluation improves total running time (up to $6.2\times$) and latency (up to $12.7\times$) compared to standard sequential Python, while performing very close (between $1.3\%$ and $18.5\%$ running time overhead) to hand-tuned manually optimized asynchronous Rust. For Tree-of-Thoughts, a prominent LLM reasoning approach, we achieve a $6.2\times$ performance improvement over the authors' own implementation.

cs.PL↗

arXiv2025

Alignment of large language models with constrained learning

We study the problem of computing an optimal large language model (LLM) policy for the constrained alignment problem, where the goal is to maximize a primary reward objective while satisfying constraints on secondary utilities. Despite the popularity of Lagrangian-based LLM policy search in constrained alignment, iterative primal-dual methods often fail to converge, and non-iterative dual-based methods do not achieve optimality in the LLM parameter space. To address these challenges, we employ Lagrangian duality to develop an iterative dual-based alignment method that alternates between updating the LLM policy via Lagrangian maximization and updating the dual variable via dual descent. In theory, we characterize the primal-dual gap between the primal value in the distribution space and the dual value in the LLM parameter space. We further quantify the optimality gap of the learned LLM policies at near-optimal dual variables with respect to both the objective and the constraint functions. These results prove that dual-based alignment methods can find an optimal constrained LLM policy, up to an LLM parametrization gap. We demonstrate the effectiveness and merits of our approach through extensive experiments conducted on the PKU-SafeRLHF and Anthropic HH-RLHF datasets.

cs.LG↗

arXiv2025

BrowserArena: Evaluating LLM Agents on Real-World Web Navigation Tasks

LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about pop-up banner closure. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.

cs.AI↗

arXiv2025

RAPID: An Efficient Reinforcement Learning Algorithm for Small Language Models

Reinforcement learning (RL) has emerged as a promising strategy for finetuning small language models (SLMs) to solve targeted tasks such as math and coding. However, RL algorithms tend to be resource-intensive, taking a significant amount of time to train. We propose RAPID, a novel RL algorithm that can substantially reduce the running time of RL. Our key insight is that RL tends to be costly due to the need to perform both inference and backpropagation during training. To maximize use of computational resources, our algorithm performs inference in large batches, and then performs off-policy policy gradient updates in mini-batches. For off-policy updates, we incorporate group advantage estimation into the policy gradient algorithm, and derive an importance weighted estimator to correct for the bias arising from off-policy learning. Our experiments demonstrate that our algorithm can reduce running time by 11%-34% on three benchmarks compared to state-of-the-art RL algorithms while maintaining similar or better accuracy.

cs.LG↗

arXiv2025

Active Learning for Neurosymbolic Program Synthesis

The goal of active learning for program synthesis is to synthesize the desired program by asking targeted questions that minimize user interaction. While prior work has explored active learning in the purely symbolic setting, such techniques are inadequate for the increasingly popular paradigm of neurosymbolic program synthesis, where the synthesized program incorporates neural components. When applied to the neurosymbolic setting, such techniques can -- and, in practice, do -- return an unintended program due to mispredictions of neural components. This paper proposes a new active learning technique that can handle the unique challenges posed by neural network mispredictions. Our approach is based upon a new evaluation strategy called constrained conformal evaluation (CCE), which accounts for neural mispredictions while taking into account user-provided feedback. Our proposed method iteratively makes CCE more precise until all remaining programs are guaranteed to be observationally equivalent. We have implemented this method in a tool called SmartLabel and experimentally evaluated it on three neurosymbolic domains. Our results demonstrate that SmartLabel identifies the ground truth program for 98% of the benchmarks, requiring under 5 rounds of user interaction on average. In contrast, prior techniques for active learning are only able to converge to the ground truth program for at most 65% of the benchmarks.

cs.PL↗

arXiv2025

Rethinking Algorithmic Fairness for Human-AI Collaboration

Existing approaches to algorithmic fairness aim to ensure equitable outcomes if human decision-makers comply perfectly with algorithmic decisions. However, perfect compliance with the algorithm is rarely a reality or even a desirable outcome in human-AI collaboration. Yet, recent studies have shown that selective compliance with fair algorithms can amplify discrimination relative to the prior human policy. As a consequence, ensuring equitable outcomes requires fundamentally different algorithmic design principles that ensure robustness to the decision-maker's (a priori unknown) compliance pattern. We define the notion of compliance-robustly fair algorithmic recommendations that are guaranteed to (weakly) improve fairness in decisions, regardless of the human's compliance pattern. We propose a simple optimization strategy to identify the best performance-improving compliance-robustly fair policy. However, we show that it may be infeasible to design algorithmic recommendations that are simultaneously fair in isolation, compliance-robustly fair, and more accurate than the human policy; thus, if our goal is to improve the equity and accuracy of human-AI collaboration, it may not be desirable to enforce traditional algorithmic fairness constraints. We illustrate the value of our approach on criminal sentencing data before and after the introduction of an algorithmic risk assessment tool in Virginia.

cs.LG↗

arXiv2025

Large Scale Multi-Task Bayesian Optimization with Large Language Models

In multi-task Bayesian optimization, the goal is to leverage experience from optimizing existing tasks to improve the efficiency of optimizing new ones. While approaches using multi-task Gaussian processes or deep kernel transfer exist, the performance improvement is marginal when scaling beyond a moderate number of tasks. We introduce a novel approach leveraging large language models (LLMs) to learn from, and improve upon, previous optimization trajectories, scaling to approximately 1500 distinct tasks. Specifically, we propose a feedback loop in which an LLM is fine-tuned on the high quality solutions to specific tasks found by Bayesian optimization (BO). This LLM is then used to generate initialization points for future BO searches for new tasks. The trajectories of these new searches provide additional training data for fine-tuning the LLM, completing the loop. We evaluate our method on two distinct domains: database query optimization and antimicrobial peptide design. Results demonstrate that our approach creates a positive feedback loop, where the LLM's generated initializations gradually improve, leading to better optimization performance. As this feedback loop continues, we find that the LLM is eventually able to generate solutions to new tasks in just a few shots that are better than the solutions produced by "from scratch" by Bayesian optimization while simultaneously requiring significantly fewer oracle calls.

cs.LG↗

arXiv2025

Stochastic Online Conformal Prediction with Semi-Bandit Feedback

Conformal prediction has emerged as an effective strategy for uncertainty quantification by modifying a model to output sets of labels instead of a single label. These prediction sets come with the guarantee that they contain the true label with high probability. However, conformal prediction typically requires a large calibration dataset of i.i.d. examples. We consider the online learning setting, where examples arrive over time, and the goal is to construct prediction sets dynamically. Departing from existing work, we assume semi-bandit feedback, where we only observe the true label if it is contained in the prediction set. For instance, consider calibrating a document retrieval model to a new domain; in this setting, a user would only be able to provide the true label if the target document is in the prediction set of retrieved documents. We propose a novel conformal prediction algorithm targeted at this setting, and prove that it obtains sublinear regret compared to the optimal conformal predictor. We evaluate our algorithm on a retrieval task, an image classification task, and an auction price-setting task, and demonstrate that it empirically achieves good performance compared to several baselines.

cs.LG↗