Large language models are increasingly deployed in multi-agent workflows. We introduce Prompt Choreography, a framework that efficiently executes LLM workflows by maintaining a dynamic, global key-value cache. Each LLM call can attend to an arbitrary, reordered subset of previously encoded messages. Parallel calls are supported. Though caching messages' encodings sometimes gives different results from re-encoding them in a new context, we show in diverse settings that fine-tuning the LLM to work with the cache can help it mimic the original results. Prompt Choreography significantly reduces per-message latency (2.0-6.2 & times; faster time-to-first-token) and achieves substantial end-to-end speedups (>2.2 & times;) in some workflows dominated by redundant computation.

Large language model-based agents are increasingly deployed as collaborators in scientific discovery yet most current work focuses on the autonomous capabilities of "AI Scientists". We argue that this overlooks the social aspects of scientific teamwork, and that studying AI Scientists as human-agent systems (HAS)–where the unit of analysis is the human-agent pair–is both underexplored and undervalued. We establish these points through literature and empirical analysis, and highlight recent incidences and studies which show that deploying agents in science without accounting for human-agent dynamics introduces near-term risks, including reduced diversity of scientific inquiry. Through analysis of real-world case studies, we show that scientists and agents can augment each other's capabilities. We call for new research that adopts the HAS lens to develop mathematical frameworks for understanding and fostering human-AI synergy in scientific discovery.

Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, 5.7 × 10^2 or 580?" This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3

A wide range of LM applications require generating text that conforms to syntactic or semantic constraints. Imposing such constraints can be naturally framed as probabilistic conditioning, but exact generation from the resulting distribution—which can differ substantially from the LM’s base distribution—is generally intractable. In this work, we develop an architecture for controlled LM generation based on sequential Monte Carlo (SMC). This SMC framework allows us to flexibly incorporate domain- and problem-specific constraints at inferencetime, and efficiently reallocate computational resources in light of new information during the course of generation. By comparing to a number of alternatives and ablations on four challenging domains—Python code generation for data science, text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with little overhead, our approach allows small open-source language models to outperform models over 8× larger, as well as closed-source, fine-tuned ones. In support of the probabilistic perspective, we show that these performance improvements are driven by better approximation to the posterior distribution. [Our system](https://github.com/probcomp/gen-parse) builds on the framework of Lew et al. (2023) and integrates with its language model probabilistic programming language, giving users a simple, programmable way to apply SMC to a broad variety of controlled generation problems.

We present a method and dataset for fine-tuning language models with preference supervision using feedback-driven improvement chains. Given a model response, an annotator provides fine-grained feedback by marking ``liked'' and ``disliked'' spans and specifying what they liked or disliked about them. The base model then rewrites the disliked spans accordingly, proceeding from left to right, forming a sequence of incremental improvements. We construct preference pairs for direct alignment from each adjacent step in the chain, enabling the model to learn from localized, targeted edits. We find that our approach outperforms direct alignment methods based on standard A/B preference ranking or full contrastive rewrites, demonstrating that structured, revision-based supervision leads to more efficient and effective preference tuning.

The dominant approach to generating from language models subject to some constraint is locally constrained decoding (LCD), incrementally sampling tokens at each time step such that the constraint is never violated. Typically, this is achieved through token masking: looping over the vocabulary and excluding non-conforming tokens. There are two important problems with this approach. (i) Evaluating the constraint on every token can be prohibitively expensive – LM vocabularies often exceed 100,000 tokens. (ii) LCD can distort the global distribution over strings, sampling tokens based only on local information, even if they lead down dead-end paths. This work introduces a new algorithm that addresses both these problems. First, to avoid evaluating a constraint on the full vocabulary at each step of generation, we propose an adaptive rejection sampling algorithm that typically requires orders of magnitude fewer constraint evaluations. Second, we show how this algorithm can be extended to produce low-variance, unbiased estimates of importance weights at a very small additional cost – estimates that can be soundly used within previously proposed sequential Monte Carlo algorithms to correct for the myopic behavior of local constraint enforcement. Through extensive empirical evaluation in text-to-SQL, molecular synthesis, goal inference, pattern matching, and JSON domains, we show that our approach is superior to state-of-the-art baselines, supporting a broader class of constraints and improving both runtime and performance. Additional theoretical and empirical analyses show that our method's runtime efficiency is driven by its dynamic use of computation, scaling with the divergence between the unconstrained and constrained LM, and as a consequence, runtime improvements are greater for better models.

Tool-using agents that act in the world need to be both useful and safe. Well-calibrated model confidences can be used to weigh the risk versus reward of potential actions, but prior work shows that many models are poorly calibrated. Inspired by interpretability literature exploring the internals of models, we propose a novel class of model-internal confidence estimators (MICE) to better assess confidence when calling tools. MICE first decodes from each intermediate layer of the language model using logit lens (nostalgebraist, 2020) and then computes similarity scores between each layer's generation and the final output. These features are fed into a learned probabilistic classifier to assess confidence in the decoded output. On the simulated trial and error (STE) tool-calling dataset using Llama3 models, we find that MICE beats or matches the baselines on smoothed expected calibration error. Using MICE confidences to determine whether to call a tool significantly improves over strong baselines on a new metric, expected tool-calling utility. Further experiments show that MICE is sample-efficient, can generalize zero-shot to unseen APIs, and results in higher tool-calling utility in scenarios with varying risk levels. Our code is open source, available at https: //github.com/microsoft/mice_for_cats.

Much algorithmic research in NLP aims to efficiently manipulate rich formal structures. An algorithm designer typically seeks to provide guarantees about their proposed algorithm -- for example, that its running time or space complexity is upper-bounded as a certain function of its input size. They may also wish to determine the necessary properties of the quantities derived by the algorithm to synthesize efficient data structures and verify type errors. In this paper, we develop a system for helping programmers to perform these types of analyses. We apply our system to a number of NLP algorithms and find that it successfully infers types, dead and redundant code, and parametric runtime and space complexity bounds.

A hallmark of intelligence is the ability to flesh out underspecified situations using "common sense." We propose to extract that common sense from large language models (LLMs), in a form that can feed into probabilistic inference. We focus our investigation on guesstimation questions such as "How much are Airbnb listings in Newark, NJ?" Formulating a sensible answer without access to data requires drawing on, and integrating, bits of common knowledge about how and may relate to other variables, such as . Our framework answers such a question by synthesizing an ad hoc probabilistic model. First we prompt an LLM to propose a set of random variables relevant to the question, followed by moment constraints on their joint distribution. We then optimize the joint distribution p within a log-linear family to maximize the overall constraint satisfaction. Our experiments show that LLMs can successfully be prompted to propose reasonable variables, and while the proposed numerical constraints can be noisy, jointly optimizing for their satisfaction reconciles them. When evaluated on probabilistic questions derived from three real-world tabular datasets, we find that our framework performs comparably to a direct prompting baseline in terms of total variation distance from the dataset distribution, and is similarly robust to noise.

Large language models (LLMs) have an impressive ability to draw on novel information supplied in their context. Yet the mechanisms underlying this contextual grounding remain unknown, especially in situations where contextual information contradicts factual knowledge stored in the parameters, which LLMs also excel at recalling. Favoring the contextual information is critical for retrieval-augmented generation methods, which enrich the context with up-to-date information, hoping that grounding can rectify outdated or noisy stored knowledge. We present a novel method to study grounding abilities using Fakepedia, a novel dataset of counterfactual texts constructed to clash with a model's internal parametric knowledge. In this study, we introduce Fakepedia, a counterfactual dataset designed to evaluate grounding abilities when the internal parametric knowledge clashes with the contextual information. We benchmark various LLMs with Fakepedia and conduct a causal mediation analysis of LLM components when answering Fakepedia queries, based on our Masked Grouped Causal Tracing (MGCT) method. Through this analysis, we identify distinct computational patterns between grounded and ungrounded responses. We finally demonstrate that distinguishing grounded from ungrounded responses is achievable through computational analysis alone. Our results, together with existing findings about factual recall mechanisms, provide a coherent narrative of how grounding and factual recall mechanisms interact within LLMs.

This paper introduces a framework for the automated evaluation of natural language texts. A manually constructed rubric describes how to assess multiple dimensions of interest. To evaluate a text, a large language model (LLM) is prompted with each rubric question and produces a distribution over potential responses. The LLM predictions often fail to agree well with human judges -- indeed, the humans do not fully agree with one another. However, the multiple LLM distributions can be $\textit{combined}$ to $\textit{predict}$ each human judge's annotations on all questions, including a summary question that assesses overall quality or relevance. LLM-Rubric accomplishes this by training a small feed-forward neural network that includes both judge-specific and judge-independent parameters. When evaluating dialogue systems in a human-AI information-seeking task, we find that LLM-Rubric with 9 questions (assessing dimensions such as naturalness, conciseness, and citation quality) predicts human judges' assessment of overall user satisfaction, on a scale of 1--4, with RMS error $< 0.5$, a $2\times$ improvement over the uncalibrated baseline.

A language model may be viewed as a Svalued stochastic process for some alphabet S. However, in some pathological situations, such a stochastic process may "leak" probability mass onto the set of infinite strings and hence is not equivalent to the conventional view of a language model as a distribution over ordinary (finite) strings. Such ill-behaved language processes are referred to as non-tight in the literature. In this work, we study conditions of tightness through the lens of stochastic processes. In particular, by regarding the EOS symbol as marking a stopping time and using results from martingale theory, we give characterizations of tightness that generalize our previous work (Du et al., 2023).

We describe a class of tasks called decision-oriented dialogues, in which AI assistants such as large language models (LMs) must collaborate with one or more humans via natural language to help them make complex decisions. We formalize three domains in which users face everyday decisions: (1) choosing an assignment of reviewers to conference papers, (2) planning a multi-step itinerary in a city, and (3) negotiating travel plans for a group of friends. In each of these settings, AI assistants and users have disparate abilities that they must combine to arrive at the best decision: Assistants can access and process large amounts of information, while users have preferences and constraints external to the system. For each task, we build a dialogue environment where agents receive a reward based on the quality of the final decision they reach. We evaluate LMs in self-play and in collaboration with humans and find that they fall short compared to human assistants, achieving much lower rewards despite engaging in longer dialogues. We highlight a number of challenges models face in decision-oriented dialogues, ranging from goal-directed behavior to reasoning and optimization, and release our environments as a testbed for future work.

We consider the problem of sampling text from an energy-based model. This arises, for example, when sampling text from a neural language model subject to soft constraints. Although the target distribution is discrete, the internal computations of the energy function (given by the language model) are differentiable, so one would like to exploit gradient information within a method such as MCMC. Alas, all previous attempts to generalize gradient-based MCMC to text sampling fail to sample correctly from the target distribution. We propose a solution, along with variants, and study its theoretical properties. Through experiments on various forms of text generation, we demonstrate that our unbiased samplers are able to generate more fluent text while better adhering to the control objectives. The same methods could be used to sample from discrete energy-based models unrelated to text.

We design probes trained on the internal representations of a transformer language model to predict its hallucinatory behavior on three grounded generation tasks. To train the probes, we annotate for span-level hallucination on both sampled (organic) and manually edited (synthetic) reference outputs. Our probes are narrowly trained and we find that they are sensitive to their training domain: they generalize poorly from one task to another or from synthetic to organic hallucinations. However, on in-domain data, they can reliably detect hallucinations at many transformer layers, achieving 95% of their peak performance as early as layer 4. Here, probing proves accurate for evaluating hallucination, outperforming several contemporary baselines and even surpassing an expert human annotator in response-level detection F1. Similarly, on span-level labeling, probes are on par or better than the expert annotator on two out of three generation tasks. Overall, we find that probing is a feasible and efficient alternative to language model hallucination evaluation when model states are available.

Large language models (LLMs) can improve their accuracy on various tasks through iteratively refining and revising their output based on feedback. We observe that these revisions can introduce errors, in which case it is better to roll back to a previous result. Further, revisions are typically homogeneous: they use the same reasoning method that produced the initial answer, which may not correct errors. To enable exploration in this space, we present SCREWS, a modular framework for reasoning with revisions. It is comprised of three main modules: Sampling, Conditional Resampling, and Selection, each consisting of sub-modules that can be hand-selected per task. We show that SCREWS not only unifies several previous approaches under a common framework, but also reveals several novel strategies for identifying improved reasoning chains. We evaluate our framework with state-of-the-art LLMs (ChatGPT and GPT-4) on a diverse set of reasoning tasks and uncover useful new reasoning strategies for each: arithmetic word problems, multi-hop question answering, and code debugging. Heterogeneous revision strategies prove to be important, as does selection between original and revised candidates.

Tools for translating natural language into code promise natural, open-ended interaction with databases, web APIs, and other software systems. However, this promise is complicated by the diversity and continual development of these systems, each with its own interface and distinct set of features. Building a new language-to-code translator, even starting with a large language model (LM), typically requires annotating a large set of natural language commands with their associated programs. In this paper, we describe ICIP (In-Context Inverse Programming), a method for bootstrapping a language-to-code system using mostly (or entirely) unlabeled programs written using a potentially unfamiliar (but human-readable) library or API. ICIP uses a pre-trained LM to assign candidate natural language descriptions to these programs, then iteratively refines the descriptions to ensure global consistency. Across nine different application domains from the Overnight and Spider benchmarks and text-davinci-003 and CodeLlama-7b-Instruct models, ICIP outperforms a number of prompting baselines. Indeed, in a “nearly unsupervised” setting with only a single annotated program and 100 unlabeled examples, it achieves up to 85% of the performance of a fully supervised system.

We introduce iterative retrieval, a novel framework that empowers retrievers to make iterative decisions through policy optimization. Finding an optimal portfolio of retrieved items is a combinatorial optimization problem, generally considered NP-hard. This approach provides a learned approximation to such a solution, meeting specific task requirements under a given family of large language models (LLMs). We propose a training procedure based on reinforcement learning, incorporating feedback from LLMs. We instantiate an iterative retriever for composing in-context learning (ICL) exemplars and apply it to various semantic parsing tasks that demand synthesized programs as outputs. By adding only 4M additional parameters for state encoding, we convert an off-the-shelf dense retriever into a stateful iterative retriever, outperforming previous methods in selecting ICL exemplars on semantic parsing datasets such as CalFlow, TreeDST, and MTOP. Additionally, the trained iterative retriever generalizes across different inference LLMs beyond the one used during training.

Tools are essential for large language models (LLMs) to acquire up-to-date information and take consequential actions in external environments. Existing work on tool-augmented LLMs primarily focuses on the broad coverage of tools and the flexibility of adding new tools. However, a critical aspect that has surprisingly been understudied is simply how accurately an LLM uses tools for which it has been trained. We find that existing LLMs, including GPT-4 and open-source LLMs specifically fine-tuned for tool use, only reach a correctness rate in the range of 30% to 60%, far from reliable use in practice. We propose a biologically inspired method for tool-augmented LLMs, simulated trial and error (STE), that orchestrates three key mechanisms for successful tool use behaviors in the biological system: trial and error, imagination, and memory. Specifically, STE leverages an LLM's 'imagination' to simulate plausible scenarios for using a tool, after which the LLM interacts with the tool to learn from its execution feedback. Both short-term and long-term memory are employed to improve the depth and breadth of the exploration, respectively. Comprehensive experiments on ToolBench show that STE substantially improves tool learning for LLMs under both in-context learning and fine-tuning settings, bringing a boost of 46.7% to Mistral-Instruct-7B and enabling it to outperform GPT-4. We also show effective continual learning of tools via a simple experience replay strategy.

Users of natural language interfaces, generally powered by Large Language Models (LLMs),often must repeat their preferences each time they make a similar request. We describe an approach to LLM-based dialogue modeling in which persistent user constraints and preferences -- collectively termed standing instructions -- as additional context for such interfaces. For example, when a user states "I'm hungry", a previously expressed preference for Persian food can be automatically added to the LLM prompt, influencing the search for relevant restaurants. We develop NLSI, a language-to-program dataset consisting of over 2.4K dialogues spanning 17 domains, where each dialogue is paired with a user profile (a set of users specific standing instructions) and corresponding structured representations (API calls). A key challenge in NLSI is to identify which subset of the standing instructions is applicable to a given dialogue. NLSI contains diverse phenomena, from simple preferences to interdependent instructions such as triggering a hotel search whenever the user is booking tickets to an event. We conduct experiments on NLSI using prompting with large language models and various retrieval approaches, achieving a maximum of 44.7% exact match on API prediction. Our results demonstrate the challenges in identifying the relevant standing instructions and their interpretation into API calls.