arXiv subjects

William W. Cohen

Publications and source records attributed to William W. Cohen.

At least 19 recordsLinked to original sources

arXiv2026

Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation

Heterogeneous AI systems composed of multiple models, architectures, harnesses, or inference-time settings can improve quality and efficiency by routing queries to the specialist who can answer most effectively at the lowest cost. Routing requires estimating each specialist's expected return, but this value estimation has a cost. Cheap estimators (e.g., embedding-based predictors) are fast but noisy, while accurate estimators (e.g., fine-tuned models with access to retrieval results or partial reasoning traces) are expensive. We formalize this tradeoff as an instance of Pandora's Box, the classical problem of optimal search with costly inspection. Under a Gaussian signal model, the resulting policies have closed-form value-of-information expressions that determine, for each specialist and input, whether refining the value estimate is worth its cost. We call the centralized policy Pandora's Router. We extend this to a decentralized setting, Pandora's Bidder, where specialists independently decide whether to invest in self-assessment before accepting an offered price to claim a query. Experiments across three domains---a standard multi-LLM benchmark, retrieval-augmented specialists, and LLMs with variable inference-time reasoning---show that Pandora's Router matches the routing quality of exhaustive estimation, while querying the expensive estimator far less often. In the decentralized setting, value-of-information reasoning improves allocative efficiency when competing estimates are accurate; when competing estimates are noisy, however, it can increase the strategic specialist's utility at the expense of others.

cs.AI↗

arXiv2026

Inducing Reasoning Primitives from Agent Traces

ReAct-style LLM agents often rediscover the same reasoning routines across problems, yet leave those routines trapped in transient scratchpads. We introduce Reasoning Primitive Induction, a single-pass method that mines successful ReAct traces, clusters recurrent reasoning moves, and converts the most frequent moves into a compact library of typed pseudo-tools. Each pseudo-tool is specified by a natural-language docstring interpreted by an LLM at invocation time, and a standard ReAct loop composes these primitives at test time. The central result is that induced libraries outperform the very agent that generated their traces: by +44pp on RuleArena NBA (30 -> 74), +30pp on MuSR team allocation (38 -> 68), and +22pp on NatPlan meeting planning (7 -> 29). Across five comparable subtasks spanning narrative deduction, rule application, and constraint-satisfaction planning, a single fixed configuration improves over zero-shot Chain-of-Thought on every subtask, matches or surpasses expert-authored decompositions, and outperforms AWM at lower average inference cost.

cs.AI↗

arXiv2026

Learning to Construct Practical Agentic Systems

Automated design and optimization of agentic LLM-based systems leads to sophisticated systems that substantially improve result quality over off-the-shelf agentic patterns. However, studies of fielded agentic systems show that production systems focus much more on issues such as simplicity, controllability, and predictability of inference costs. In this paper we propose principled approaches to designing and optimizing practical agentic systems. We describe an agent framework that enables designers to enforce modularity in agentic systems, by defining "pseudo-tools" that call LLMs recursively on a restricted context. Using this framework we hand-engineer agents for a diverse set of tasks, and show that relative to dynamically-planned workflows, hand-constructed fixed workflows are generally cheaper and more accurate. We then propose novel learning methods for the agentic components required by this framework, namely pseudo-tools and fixed workflows. These learning methods generally outperform hand-engineered agents. We also exploit the modularity of the framework to apply multi-objective optimization methods to jointly optimize cost and response quality and blend the results of multiple learning systems.

cs.LG↗

arXiv2026

Multiple-Prediction-Powered Inference

Statistical estimation often involves tradeoffs between expensive, high-quality measurements and a variety of lower-quality proxies. We introduce Multiple-Prediction-Powered Inference (MultiPPI): a general framework for constructing statistically efficient estimates by optimally allocating resources across these diverse data sources. This work provides theoretical guarantees about the minimax optimality, finite-sample performance, and asymptotic normality of the MultiPPI estimator. Through experiments across three diverse large language model (LLM) evaluation scenarios, we show that MultiPPI consistently achieves lower estimation error than existing baselines. This advantage stems from its budget-adaptive allocation strategy, which strategically combines subsets of models by learning their complex cost and correlation structures.

math.ST↗

arXiv2026

Localizing and Correcting Errors for LLM-based Planners

Large language models (LLMs) have demonstrated strong reasoning capabilities on math and coding, but frequently fail on symbolic classical planning tasks. Our studies, as well as prior work, show that LLM-generated plans routinely violate domain constraints given in their instructions (e.g., walking through walls). To address this failure, we propose iteratively augmenting instructions with Localized In-Context Learning (L-ICL) demonstrations: targeted corrections for specific failing steps. Specifically, L-ICL identifies the first constraint violation in a trace and injects a minimal input-output example giving the correct behavior for the failing step. Our proposed technique of L-ICL is much effective than explicit instructions or traditional ICL, which adds complete problem-solving trajectories, and many other baselines. For example, on an 8x8 gridworld, L-ICL produces valid plans 89% of the time with only 60 training examples, compared to 59% for the best baseline, an increase of 30%. L-ICL also shows dramatic improvements in other domains (gridworld navigation, mazes, Sokoban, and BlocksWorld), and on several LLM architectures.

cs.AI↗

arXiv2025

Semi-structured LLM Reasoners Can Be Rigorously Audited

Although Large Language Models (LLMs) have become capable reasoners, the problem of faithfulness persists: their reasoning can contain errors and omissions that are difficult to detect and that may obscure biases in model outputs. To address this issue, we introduce Semi-Structured Reasoning Models (SSRMs), which are trained to produce semi-structured representations of reasoning. SSRMs generate reasoning traces in a non-executable Pythonic syntax that names each reasoning step and marks its inputs and outputs. This structure allows SSRM traces to be automatically audited to identify reasoning flaws. We evaluate three types of audits: hand-crafted structured reasoning audits, written in a domain-specific language (DSL) implemented in Python; LLM-generated structured reasoning audits; and learned typicality audits, which apply probabilistic models over reasoning traces. We show that all of these methods can be used to effectively flag probable reasoning errors. Importantly, the auditability of SSRMs does not appear to compromise overall accuracy: in evaluation on twelve benchmarks and two model families, SSRMs demonstrate strong performance and generalizability relative to other models of comparable size.

cs.CL↗

arXiv2025

CrossWordBench: Evaluating the Reasoning Capabilities of LLMs and LVLMs with Controllable Puzzle Generation

Existing reasoning evaluation frameworks for Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) predominantly assess either text-based reasoning or vision-language understanding capabilities, with limited dynamic interplay between textual and visual constraints. To address this limitation, we introduce CrossWordBench, a benchmark designed to evaluate the reasoning capabilities of both LLMs and LVLMs through the medium of crossword puzzles -- a task requiring multimodal adherence to semantic constraints from text-based clues and intersectional constraints from visual grid structures. CrossWordBench leverages a controllable puzzle generation framework that produces puzzles in two formats (text and image), supports adjustable difficulty through prefill ratio control, and offers different evaluation strategies, ranging from direct puzzle solving to interactive modes. Our extensive evaluation of over 20 models reveals that reasoning LLMs substantially outperform non-reasoning models by effectively leveraging crossing-letter constraints. We further demonstrate that LVLMs struggle with the task, showing a strong correlation between their puzzle-solving performance and grid-parsing accuracy. Our findings highlight limitations of the reasoning capabilities of current LLMs and LVLMs, and provide an effective approach for creating multimodal constrained tasks for future evaluations.

cs.CL↗

arXiv2024

Watch Your Steps: Observable and Modular Chains of Thought

We propose a variant of chain of thought (CoT) prompting called Program Trace Prompting that makes explanations more observable while preserving the power, generality and flexibility of CoT. In our approach, few-shot CoT demonstrations are wrapped in a formal syntax based on Python, and each prompt: identifies and names steps; defines the input/output behavior of steps; and replaces CoT explanations of in-context examples with chains of these formalized steps on the same examples. Program Trace Prompting is applicable to many tasks, achieving strong results on the 23 diverse tasks in the BIG-Bench Hard benchmark. More importantly, by instrumenting explanations in this way, we enable new types of analysis. In particular, we identify "non-local errors" (which correspond to incorrectly learning the reasoning method illustrated in the demonstrations) as an unaddressed issue in CoT learning, and we present methods for verifying the modularity of steps in a CoT explanation.

cs.CL↗

arXiv2024

VLM Agents Generate Their Own Memories: Distilling Experience into Embodied Programs of Thought

Large-scale generative language and vision-language models (LLMs and VLMs) excel in few-shot learning but require high-quality demonstrations. We propose In-Context Abstraction Learning (ICAL), enabling VLM agents to transform suboptimal trajectories into high-quality training data through self-reflection and human feedback. Given imperfect task demonstrations, a VLM abstracts trajectories into generalized strategies and action annotations by correcting inefficiencies and annotating cognitive abstractions: causal relationships, object state changes, temporal subgoals, and task-relevant visual elements. These annotations are iteratively refined through human feedback during execution in similar environments. The resulting examples significantly improve decision-making when used for retrieval-augmented generation or fine-tuning. As the agent's example library grows, it becomes more efficient at abstracting new examples, requiring less human feedback and fewer environment interactions. ICAL achieves state-of-the-art results across multiple benchmarks. In TEACh dialogue-based instruction following, combining fine-tuning and retrieval on ICAL examples outperforms raw human demonstrations and expert examples by 17.5% in goal-condition success. In VisualWebArena, retrieval-augmented GPT-4V with ICAL improves task success 1.6x, while fine-tuned Qwen2-VL achieves 2.8x improvement over the base model. In Ego4D action forecasting, we surpass few-shot GPT-4V and remain competitive with supervised models. Our approach scales 2x better than raw demonstrations and significantly reduces manual prompt engineering requirements.

cs.CV↗

arXiv2024

Stratified Prediction-Powered Inference for Hybrid Language Model Evaluation

Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data. PPI achieves this by combining small amounts of human-labeled data with larger amounts of data labeled by a reasonably accurate -- but potentially biased -- automatic system, in a way that results in tighter confidence intervals for certain parameters of interest (e.g., the mean performance of a language model). In this paper, we propose a method called Stratified Prediction-Powered Inference (StratPPI), in which we show that the basic PPI estimates can be considerably improved by employing simple data stratification strategies. Without making any assumptions on the underlying automatic labeling system or data distribution, we derive an algorithm for computing provably valid confidence intervals for population parameters (such as averages) that is based on stratified sampling. In particular, we show both theoretically and empirically that, with appropriate choices of stratification and sample allocation, our approach can provide substantially tighter confidence intervals than unstratified approaches. Specifically, StratPPI is expected to improve in cases where the performance of the autorater varies across different conditional distributions of the target data.

cs.LG↗

arXiv2024

Bayesian Prediction-Powered Inference

Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data. Specifically, PPI methods provide tighter confidence intervals by combining small amounts of human-labeled data with larger amounts of data labeled by a reasonably accurate, but potentially biased, automatic system. We propose a framework for PPI based on Bayesian inference that allows researchers to develop new task-appropriate PPI methods easily. Exploiting the ease with which we can design new metrics, we propose improved PPI methods for several importantcases, such as autoraters that give discrete responses (e.g., prompted LLM ``judges'') and autoraters with scores that have a non-linear relationship to human scores.

cs.LG↗

arXiv2023

Characterizing Tradeoffs in Language Model Decoding with Informational Interpretations

We propose a theoretical framework for formulating language model decoder algorithms with dynamic programming and information theory. With dynamic programming, we lift the design of decoder algorithms from the logit space to the action-state value function space, and show that the decoding algorithms are consequences of optimizing the action-state value functions. Each component in the action-state value function space has an information theoretical interpretation. With the lifting and interpretation, it becomes evident what the decoder algorithm is optimized for, and hence facilitating the arbitration of the tradeoffs in sensibleness, diversity, and attribution.

cs.CL↗

arXiv2023

Trusted Source Alignment in Large Language Models

Large language models (LLMs) are trained on web-scale corpora that inevitably include contradictory factual information from sources of varying reliability. In this paper, we propose measuring an LLM property called trusted source alignment (TSA): the model's propensity to align with content produced by trusted publishers in the face of uncertainty or controversy. We present FactCheckQA, a TSA evaluation dataset based on a corpus of fact checking articles. We describe a simple protocol for evaluating TSA and offer a detailed analysis of design considerations including response extraction, claim contextualization, and bias in prompt formulation. Applying the protocol to PaLM-2, we find that as we scale up the model size, the model performance on FactCheckQA improves from near-random to up to 80% balanced accuracy in aligning with trusted sources.

cs.CL↗

arXiv2023

SEMQA: Semi-Extractive Multi-Source Question Answering

Recently proposed long-form question answering (QA) systems, supported by large language models (LLMs), have shown promising capabilities. Yet, attributing and verifying their generated abstractive answers can be difficult, and automatically evaluating their accuracy remains an ongoing challenge. In this work, we introduce a new QA task for answering multi-answer questions by summarizing multiple diverse sources in a semi-extractive fashion. Specifically, Semi-extractive Multi-source QA (SEMQA) requires models to output a comprehensive answer, while mixing factual quoted spans -- copied verbatim from given input sources -- and non-factual free-text connectors that glue these spans together into a single cohesive passage. This setting bridges the gap between the outputs of well-grounded but constrained extractive QA systems and more fluent but harder to attribute fully abstractive answers. Particularly, it enables a new mode for language models that leverages their advanced language generation capabilities, while also producing fine in-line attributions by-design that are easy to verify, interpret, and evaluate. To study this task, we create the first dataset of this kind, QuoteSum, with human-written semi-extractive answers to natural and generated questions, and define text-based evaluation metrics. Experimenting with several LLMs in various settings, we find this task to be surprisingly challenging, demonstrating the importance of QuoteSum for developing and studying such consolidation capabilities.

cs.CL↗

arXiv2023

MEMORY-VQ: Compression for Tractable Internet-Scale Memory

Retrieval augmentation is a powerful but expensive method to make language models more knowledgeable about the world. Memory-based methods like LUMEN pre-compute token representations for retrieved passages to drastically speed up inference. However, memory also leads to much greater storage requirements from storing pre-computed representations. We propose MEMORY-VQ, a new method to reduce storage requirements of memory-augmented models without sacrificing performance. Our method uses a vector quantization variational autoencoder (VQ-VAE) to compress token representations. We apply MEMORY-VQ to the LUMEN model to obtain LUMEN-VQ, a memory model that achieves a 16x compression rate with comparable performance on the KILT benchmark. LUMEN-VQ enables practical retrieval augmentation even for extremely large retrieval corpora.

cs.CL↗

arXiv2023

Answering Ambiguous Questions with a Database of Questions, Answers, and Revisions

Many open-domain questions are under-specified and thus have multiple possible answers, each of which is correct under a different interpretation of the question. Answering such ambiguous questions is challenging, as it requires retrieving and then reasoning about diverse information from multiple passages. We present a new state-of-the-art for answering ambiguous questions that exploits a database of unambiguous questions generated from Wikipedia. On the challenging ASQA benchmark, which requires generating long-form answers that summarize the multiple answers to an ambiguous question, our method improves performance by 15% (relative improvement) on recall measures and 10% on measures which evaluate disambiguating questions from predicted outputs. Retrieving from the database of generated questions also gives large improvements in diverse passage retrieval (by matching user questions q to passages p indirectly, via questions q' generated from p).

cs.CL↗

arXiv2023

GLIMMER: generalized late-interaction memory reranker

Memory-augmentation is a powerful approach for efficiently incorporating external information into language models, but leads to reduced performance relative to retrieving text. Recent work introduced LUMEN, a memory-retrieval hybrid that partially pre-computes memory and updates memory representations on the fly with a smaller live encoder. We propose GLIMMER, which improves on this approach through 1) exploiting free access to the powerful memory representations by applying a shallow reranker on top of memory to drastically improve retrieval quality at low cost, and 2) incorporating multi-task training to learn a general and higher quality memory and live encoder. GLIMMER achieves strong gains in performance at faster speeds compared to LUMEN and FiD on the KILT benchmark of knowledge-intensive tasks.

cs.CL↗

arXiv2023

Subject-driven Text-to-Image Generation via Apprenticeship Learning

Recent text-to-image generation models like DreamBooth have made remarkable progress in generating highly customized images of a target subject, by fine-tuning an ``expert model'' for a given subject from a few examples. However, this process is expensive, since a new expert model must be learned for each subject. In this paper, we present SuTI, a Subject-driven Text-to-Image generator that replaces subject-specific fine tuning with in-context learning. Given a few demonstrations of a new subject, SuTI can instantly generate novel renditions of the subject in different scenes, without any subject-specific optimization. SuTI is powered by apprenticeship learning, where a single apprentice model is learned from data generated by a massive number of subject-specific expert models. Specifically, we mine millions of image clusters from the Internet, each centered around a specific visual subject. We adopt these clusters to train a massive number of expert models, each specializing in a different subject. The apprentice model SuTI then learns to imitate the behavior of these fine-tuned experts. SuTI can generate high-quality and customized subject-specific images 20x faster than optimization-based SoTA methods. On the challenging DreamBench and DreamBench-v2, our human evaluation shows that SuTI significantly outperforms existing models like InstructPix2Pix, Textual Inversion, Imagic, Prompt2Prompt, Re-Imagen and DreamBooth, especially on the subject and text alignment aspects.

cs.CV↗