Searcharxiv⌕ Search

arXiv subjects

Scott Niekum

Publications and source records attributed to Scott Niekum.

At least 19 recordsLinked to original sources
arXiv2026

Training Agents to Fail Predictably: Fine-Tuning for Forecastable Tail Risk

Estimating how often an ML model will fail at deployment scale is central to pre-deployment safety assessment, but a feasible evaluation set is rarely large enough to observe the failures that matter. Jones et al. (2025) address this by extrapolating from the largest k failure scores in an evaluation set to predict deployment-scale failure rates. We give a finite-k decomposition of this estimator's forecast error and show that it has a built-in bias toward over-prediction in the typical case, which is the safety-favorable direction. This bias is offset when the evaluation set misses a rare high-failure mode that the deployment set contains, leaving the forecast to under-predict at deployment scale. We propose a fine-tuning objective, the forecastability loss, that addresses this failure mode. In two proof-of-concept experiments, a language-model password game and an RL gridworld, fine-tuning substantially reduces held-out forecast error while preserving primary-task capability and achieving safety similar to that of supervised baselines.

cs.LG↗
arXiv2026

Towards Continual Learning: Mitigating Forgetting With Subspace Tuning

Large language models (LLMs) acquire new skills through fine-tuning, but it can destroy capabilities they already have: catastrophic forgetting remains a central obstacle to continual learning in LLMs. Recent works connect forgetting to how far, and along which directions, fine-tuning moves the weights. On-policy methods such as reinforcement learning and on-policy distillation avoid updating the principal singular directions of weight matrices and thus forget less. In contrast, SFT concentrates updates along these directions and is susceptible to catastrophic forgetting. In this work, we introduce Subspace Tuning (SST), which restricts each weight update to the bottom singular subspace, eliminating changes along the principal directions. SST is optimizer-agnostic and composes well with fine-tuning recipes including SFT and self-distillation. SST combines tight subspace protection, dense gradient estimation, and a simple refresh mechanism, enabling stronger learning than prior orthogonal fine-tuning methods. With these mechanisms, protecting as much as 99\% of the weight subspace substantially improves retention while preserving strong target-task performance. This behavior strengthens with model and data size and nearly eliminates forgetting in the large-model regime. SST also keeps the model close to its base distribution, inducing $2$--$4\times$ less KL drift than the baseline recipes. Overall, SST advances the learning--forgetting Pareto frontier, making it a practical default for fine-tuning pre-trained language models.

cs.LG↗
arXiv2026

Hierarchical Experimentalist Agents

Large language models (LLMs) are increasingly used to act in the real world and augment human decision-making, yet most systems rely on parametric knowledge acquired through imitation, optionally improved by post-training, retrieval, or search over fixed data. This paradigm struggles in novel domains or on complex queries that require information unavailable from prior knowledge alone; knowing the laws of physics, for example, does not by itself enable reliable reasoning or long-horizon actions in a complex physical system. We argue that agents need active experimentation: the ability to gather targeted query-specific evidence, discover general principles of unseen environments, and convert interaction experience into reusable skills. We introduce Hierarchical Experimentalist Agents (HExA), an in-context, experiment-centric self-improvement framework that (1) iteratively designs and refines query-relevant experiments, (2) incrementally distills experience into reusable and composable skills that accelerate experimentation within and across tasks, and (3) integrates experimental evidence to act or answer queries. HExA can work with any LLM including frontier black-box models, and requires no external supervision, oracle solutions, or offline data. We evaluate HExA on multi-step reasoning, interactive partially observable games, and physics law discovery tasks. To further evaluate when active experimentation is necessary, we introduce Interphyre, a set of hard procedural physics tasks where we enable tool-calling and intervention APIs. Current frontier LLMs including GPT-6 Astra fail in these settings: on the most challenging Interphyre tasks, we show that HExA can enable such models to near 100% success. Moreover, HExA improves token efficiency compared to naive agentic and memory baselines and enables generalization via hierarchical skill transfer from easier to harder tasks.

cs.AI↗
arXiv2026

Convex-Concave Reinforcement Learning

Policy learning drives many of the most consequential and heavily-invested applications of reinforcement learning today. Yet the core optimization problem it rests on (maximizing expected return) is notoriously non-convex, even under a direct policy parameterization, and the field has largely responded by avoiding it: optimizing convex surrogate approximations of the return under trust-region constraints (NPG, TRPO, PPO, AWR). We show that this seemingly unstructured problem is not actually structureless. In log-density-ratio coordinates $y := \log[π/π_n]$, the exact per-iteration objective, computable via per-decision importance sampling (PDIS), is a difference-of-convex-constrained difference-of-convex (DC-constrained DC) program. This structure lets us move beyond surrogate approximations: it recovers CPI, NPG, TRPO, and AWR as special cases along interpretable axes, and it opens a multi-step axis $k$ that couples consecutive decisions. We solve the per-iteration program with sequential convex programming (SCP), the standard solver for difference-of-convex problems, and give convergence guarantees under mild conditions, bridging the difference-of-convex optimization and RL literatures. Empirically, multi-step Convex-Concave RL (CCRL) wins on diagnostic MDPs where credit must propagate across a horizon (its advantage growing with the dependency length), is competitive with a tuned PPO on classic control, and on a realistic, stochastic, mid-horizon healthcare domain converges markedly faster than tuned PPO to the same near-optimal survival, with an 11.3% higher area under the training curve.

cs.LG↗
arXiv2026

Safe Inference-Time Alignment via Lagrangian Reward Augmentation

Inference-time alignment steers a frozen language model during decoding using auxiliary reward signals, avoiding the cost of repeated weight updates. However, existing inference-time alignment methods typically optimize a single scalar score, so explicit safety constraints must either be ignored or encoded through manually tuned penalties. We propose Lagrangian Reward Augmentation (LARA), a general inference-time alignment framework under safety constraints. Starting from a KL-regularized constrained objective with a reward model and a cost model, LARA dualizes the constraint and reduces the optimization problem to a one-dimensional convex problem over a nonnegative dual variable. Estimated on a small calibration set, this dual variable defines an augmented reward that can be used as a drop-in scoring signal within existing inference-time alignment methods. For sequence-level sampling methods, such as Best-of-N reranking, the calibrated dual variable corresponds to the solution of the expected-cost constrained problem. For token-level reward-guided decoding methods, the same construction yields a principled dual-calibrated heuristic rather than an exact constrained-policy guarantee. We evaluate LARA on both sequence-level and token-level inference-time alignment methods, and find that LARA improves the helpfulness-harmlessness tradeoff, with Best-of-N achieving the best performance among inference-time methods, approaching finetuning-based direct alignment baselines.

cs.LG↗
arXiv2026

Regularized Latent Dynamics Prediction is a Strong Baseline For Behavioral Foundation Models

Behavioral Foundation Models (BFMs) produce agents with the capability to adapt to any unknown reward or task. These methods, however, are only able to produce near-optimal policies for the reward functions that are in the span of some pre-existing state features, making the choice of state features crucial to the expressivity of the BFM. As a result, BFMs are trained using a variety of complex objectives and require sufficient dataset coverage, to train task-useful spanning features. In this work, we examine the question: are these complex representation learning objectives necessary for zero-shot RL? Specifically, we revisit the objective of self-supervised next-state prediction in latent space for state feature learning, but observe that such an objective alone is prone to increasing state-feature similarity, and subsequently reducing span. We propose an approach, Regularized Latent Dynamics Prediction (RLDP), that adds a simple orthogonality regularization to maintain feature diversity and can match or surpass state-of-the-art complex representation learning methods for zero-shot RL. Furthermore, we empirically show that prior approaches perform poorly in low-coverage scenarios where RLDP still succeeds.

cs.AI↗
arXiv2026

Supervised Reward Inference

Existing approaches to reward inference typically assume that humans provide demonstrations according to specific behavior models. However, humans often indicate their goals through a wide range of behaviors, from actions that are suboptimal due to poor planning or execution to behaviors intended to communicate goals rather than achieve them. One existing solution for inferring rewards from such behavior $\unicode{x2013}$ provided it is drawn from the same distribution at training and deployment $\unicode{x2013}$ is to construct a dataset of behavior paired with known rewards, and to learn the mapping from behavior to rewards; however, prior methods in this family face notable limitations, such as restrictions to tabular settings. Given such a dataset, we propose instead that supervised learning offers a parsimonious yet powerful solution, which we term Supervised Reward Inference (SRI). Theoretically, we prove that SRI is asymptotically Bayes-optimal under standard assumptions. Empirically, SRI achieves near-ceiling performance on a prior benchmark for reward inference from suboptimal behavior, while on Meta-World robotics tasks, it infers rewards from even arbitrarily suboptimal demonstrations as accurately as those demonstrations allow. Finally, we demonstrate our framework's universality with straightforward generalizations to action- and goal-prediction.

cs.LG↗
arXiv2026

A Descriptive and Normative Theory of Human Beliefs in RLHF

Human preferences in RLHF are typically modeled as a function of the human's reward function or corresponding optimal state-action values. In this work, we propose that human beliefs about the capabilities of the agent being trained also play a key role in preference generation. We examine two questions related to this hypothesis, one descriptive and one normative, respectively: Do human labelers' beliefs about agent capabilities affect the preferences that they provide? And what is the ideal set of beliefs about an agent -- and resulting preferences -- for humans to have? We propose a new preference model that incorporates human beliefs and provide a normative theory that bounds the error on the final learned policy based on the mismatch between the human's beliefs and an idealized set of beliefs. We then confirm via a human study that beliefs about agent capabilities do, in fact, significantly affect preferences and can be influenced through simple interventions. Additionally, we empirically show through synthetic experiments that it is often suboptimal for human preference labelers to assume agent optimality. Collectively, these results theoretically and empirically demonstrate how reducing the mismatch between human beliefs and agent capabilities can lead to more performant RLHF and point toward new best practices for RLHF practitioners.

cs.AI↗
arXiv2026

Adaptive Margin RLHF via Preference over Preferences

Margin-based optimization is fundamental to improving generalization and robustness in classification tasks. In the context of reward model learning from preferences within Reinforcement Learning from Human Feedback (RLHF), existing methods typically rely on no margins, fixed margins, or margins that are simplistic functions of preference ratings. However, such formulations often fail to account for the varying strengths of different preferences or they rely on noisy margin information derived from preference ratings. Furthermore, many existing methods that use adaptive margins assume access to accurate preference scores, which can be difficult for humans to provide reliably. We propose leveraging preferences over preferences, that is, annotations indicating which of two preferences reflects a stronger distinction, to infer adaptive margins on a per-datapoint basis. Such preference-over-preference annotations are general and can be incorporated into both standard RLHF reward modeling objectives and direct alignment losses. As a concrete instantiation, we introduce DPO-PoP, an extension to Direct Preference Optimization (DPO) that incorporates adaptive margins from preference-over-preference supervision, enabling improved discriminative and generative performance. Additionally, we show a tradeoff between discriminative and generative performance and propose two sampling strategies for gathering preference-over-preference labels to navigate it.

cs.LG↗
arXiv2026

Safe RLHF Beyond Expectation: Stochastic Dominance for Universal Spectral Risk Control

Safe Reinforcement Learning from Human Feedback (RLHF) typically enforces safety through expected cost constraints, but the expectation captures only a single statistic of the cost distribution and fails to account for distributional uncertainty, particularly under heavy tails or rare catastrophic events. This limitation is problematic when robustness and risk sensitivity are critical. Stochastic dominance offers a principled alternative by comparing entire cost distributions rather than just their averages, enabling direct control over tail risks and potential out-of-distribution failures that expectation-based constraints may overlook. In this work, we propose Risk-sensitive Alignment via Dominance (RAD), a novel alignment framework that replaces scalar expected cost constraints with First-Order Stochastic Dominance (FSD) constraints. We operationalize this constraint by comparing the target policy's cost distribution to that of a reference policy within an Optimal Transport (OT) framework, using entropic regularization and Sinkhorn iterations to obtain a differentiable and computationally efficient objective for stable end-to-end optimization. Furthermore, we introduce quantile-weighted FSD constraints and show that weighted FSD universally controls a broad class of Spectral Risk Measures (SRMs), so that improvements under weighted dominance imply guaranteed improvements in the corresponding spectral risk. This provides a principled mechanism for tuning a model's risk profile via the quantile weighting function. Empirical results demonstrate that RAD improves harmlessness over baselines while remaining competitive in helpfulness, and exhibits greater robustness on out-of-distribution harmlessness evaluations.

cs.LG↗
arXiv2026

The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching

Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.

cs.LG↗
arXiv2026

DeepFleet: Multi-Agent Foundation Models for Mobile Robots

We introduce DeepFleet, a suite of foundation models designed to support coordination and planning for large-scale mobile robot fleets. These models are trained on fleet movement data, including robot positions, goals, and interactions, from hundreds of thousands of robots in Amazon warehouses worldwide. DeepFleet consists of four architectures that each embody a distinct inductive bias and collectively explore key points in the design space for multi-agent foundation models: the robot-centric (RC) model is an autoregressive decision transformer operating on neighborhoods of individual robots; the robot-floor (RF) model uses a transformer with cross-attention between robots and the warehouse floor; the image-floor (IF) model applies convolutional encoding to a multi-channel image representation of the full fleet; and the graph-floor (GF) model combines temporal attention with graph neural networks for spatial relationships. In this paper, we describe these models and present our evaluation of the impact of these design choices on prediction task performance. We find that the robot-centric and graph-floor models, which both use asynchronous robot state updates and incorporate the localized structure of robot interactions, show the most promise. We also present experiments that show that these two models can make effective use of larger warehouses operation datasets as the models are scaled up.

cs.RO↗
arXiv2026

Evaluation-Aware Reinforcement Learning

Policy evaluation is a core component of many reinforcement learning (RL) algorithms and a critical tool for ensuring safe deployment of RL policies. However, existing policy evaluation methods often suffer from high variance or bias. To address these issues, we introduce Evaluation-Aware Reinforcement Learning (EvA-RL), a general policy learning framework that considers evaluation accuracy at train-time, as opposed to standard post-hoc policy evaluation methods. Specifically, EvA-RL directly optimizes policies for efficient and accurate evaluation, in addition to being performant. We provide an instantiation of EvA-RL and demonstrate through a combination of theoretical analysis and empirical results that EvA-RL effectively trades off between evaluation accuracy and expected return. Finally, we show that the evaluation-aware policy and the evaluation mechanism itself can be co-learned to mitigate this tradeoff, providing the evaluation benefits without significantly sacrificing policy performance. This work opens a new line of research that elevates reliable evaluation to a first-class principle in reinforcement learning.

cs.AI↗
arXiv2025

RLZero: Direct Policy Inference from Language Without In-Domain Supervision

The reward hypothesis states that all goals and purposes can be understood as the maximization of a received scalar reward signal. However, in practice, defining such a reward signal is notoriously difficult, as humans are often unable to predict the optimal behavior corresponding to a reward function. Natural language offers an intuitive alternative for instructing reinforcement learning (RL) agents, yet previous language-conditioned approaches either require costly supervision or test-time training given a language instruction. In this work, we present a new approach that uses a pretrained RL agent trained using only unlabeled, offline interactions--without task-specific supervision or labeled trajectories--to get zero-shot test-time policy inference from arbitrary natural language instructions. We introduce a framework comprising three steps: imagine, project, and imitate. First, the agent imagines a sequence of observations corresponding to the provided language description using video generative models. Next, these imagined observations are projected into the target environment domain. Finally, an agent pretrained in the target environment with unsupervised RL instantly imitates the projected observation sequence through a closed-form solution. To the best of our knowledge, our method, RLZero, is the first approach to show direct language-to-behavior generation abilities on a variety of tasks and environments without any in-domain supervision. We further show that components of RLZero can be used to generate policies zero-shot from cross-embodied videos, such as those available on YouTube, even for complex embodiments like humanoids.

cs.AI↗
arXiv2025

Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluation

Large Language Models (LLMs) are widely used as proxies for human labelers in both training (Reinforcement Learning from AI Feedback) and large-scale response evaluation (LLM-as-a-judge). Alignment and evaluation are critical components in the development of reliable LLMs, and the choice of feedback protocol plays a central role in both but remains understudied. In this work, we show that the choice of feedback protocol for evaluation (absolute scores versus relative preferences) can significantly affect evaluation reliability and induce systematic biases. In the context of LLM-as-a-judge evaluation, we show that pairwise protocols are more vulnerable to distracted evaluation. Generator models can exploit spurious attributes (or distractor features) favored by the LLM judge, resulting in inflated scores for lower-quality outputs. We find that absolute scoring is more robust to such manipulation, producing judgments that better reflect response quality and are less influenced by distractor features. Our results demonstrate that generator models can flip preferences by embedding distractor features, skewing LLM-as-a-judge comparisons and leading to inaccurate conclusions about model quality in benchmark evaluations. Pairwise preferences flip in about 35% of the cases, compared to only 9% for absolute scores. We offer recommendations for choosing feedback protocols based on dataset characteristics and evaluation objectives.

cs.LG↗
arXiv2025

Reinforcement Learning from Human Feedback with High-Confidence Safety Constraints

Existing approaches to language model alignment often treat safety as a tradeoff against helpfulness, which can lead to unacceptable responses in sensitive domains. To ensure reliable performance in such settings, we propose High-Confidence Safe Reinforcement Learning from Human Feedback (HC-RLHF), a method that provides high-confidence safety guarantees while maximizing helpfulness. Similar to previous methods, HC-RLHF explicitly decouples human preferences into helpfulness and harmlessness (safety), which are learned by training a reward model and a cost model, respectively. It then employs a two-step process to find safe solutions. In the first step, it optimizes the reward function under an intentionally pessimistic version of the cost constraint. In the second step, the trained model undergoes a safety test to verify whether its performance stays within an upper-confidence bound of the actual cost constraint. We provide a theoretical analysis of HC-RLHF, including proof that it will not return an unsafe solution with a probability greater than a user-specified threshold. For our empirical analysis, we apply HC-RLHF to align three different language models (Qwen2-1.5B, Qwen2.5-3B, and LLaMa3.2-3B) with human preferences. Our results demonstrate that HC-RLHF produces safe models with high probability and can improve harmlessness and helpfulness compared to previous methods.

cs.LG↗
arXiv2025

Null Counterfactual Factor Interactions for Goal-Conditioned Reinforcement Learning

Hindsight relabeling is a powerful tool for overcoming sparsity in goal-conditioned reinforcement learning (GCRL), especially in certain domains such as navigation and locomotion. However, hindsight relabeling can struggle in object-centric domains. For example, suppose that the goal space consists of a robotic arm pushing a particular target block to a goal location. In this case, hindsight relabeling will give high rewards to any trajectory that does not interact with the block. However, these behaviors are only useful when the object is already at the goal -- an extremely rare case in practice. A dataset dominated by these kinds of trajectories can complicate learning and lead to failures. In object-centric domains, one key intuition is that meaningful trajectories are often characterized by object-object interactions such as pushing the block with the gripper. To leverage this intuition, we introduce Hindsight Relabeling using Interactions (HInt), which combines interactions with hindsight relabeling to improve the sample efficiency of downstream RL. However because interactions do not have a consensus statistical definition tractable for downstream GCRL, we propose a definition of interactions based on the concept of null counterfactual: a cause object is interacting with a target object if, in a world where the cause object did not exist, the target object would have different transition dynamics. We leverage this definition to infer interactions in Null Counterfactual Interaction Inference (NCII), which uses a "nulling'' operation with a learned model to infer interactions. NCII is able to achieve significantly improved interaction inference accuracy in both simple linear dynamics domains and dynamic robotic domains in Robosuite, Robot Air Hockey, and Franka Kitchen and HInt improves sample efficiency by up to 4x.

cs.LG↗
arXiv2025

An Optimal Discriminator Weighted Imitation Perspective for Reinforcement Learning

We introduce Iterative Dual Reinforcement Learning (IDRL), a new method that takes an optimal discriminator-weighted imitation view of solving RL. Our method is motivated by a simple experiment in which we find training a discriminator using the offline dataset plus an additional expert dataset and then performing discriminator-weighted behavior cloning gives strong results on various types of datasets. That optimal discriminator weight is quite similar to the learned visitation distribution ratio in Dual-RL, however, we find that current Dual-RL methods do not correctly estimate that ratio. In IDRL, we propose a correction method to iteratively approach the optimal visitation distribution ratio in the offline dataset given no addtional expert dataset. During each iteration, IDRL removes zero-weight suboptimal transitions using the learned ratio from the previous iteration and runs Dual-RL on the remaining subdataset. This can be seen as replacing the behavior visitation distribution with the optimized visitation distribution from the previous iteration, which theoretically gives a curriculum of improved visitation distribution ratios that are closer to the optimal discriminator weight. We verify the effectiveness of IDRL on various kinds of offline datasets, including D4RL datasets and more realistic corrupted demonstrations. IDRL beats strong Primal-RL and Dual-RL baselines in terms of both performance and stability, on all datasets.

cs.LG↗