Language models (LMs) are trained on billions of tokens in an attempt to recover the true language distribution. Still, vanilla random sampling from LMs yields low quality generations. Decoding algorithms attempt to restrict the LM distribution to a set of high-probability continuations, but rely on greedy heuristics that introduce myopic distortions, yielding sentences that are homogeneous, repetitive and incoherent. In this paper, we introduce EPIC, a hyperparameter-free decoding approach that incorporates the entropy of future trajectories into LM decoding. EPIC explicitly regulates the amount of uncertainty expressed at every step of generation, aligning the sampling distribution's entropy to the aleatoric (data) uncertainty. Through Entropy-Aware Lazy Gumbel-Max sampling, EPIC manages to be exact, while also being efficient, requiring only a sublinear number of entropy evaluations per step. Unlike current baselines, EPIC yields sampling distributions that are empirically well-aligned with the entropy of the underlying data distribution. Across creative writing and summarization tasks, EPIC consistently improves LM-as-judge preference win-rates over widely used decoding strategies. These preference gains are complemented by automatic metrics, showing that EPIC produces more diverse generations and more faithful summaries. We also evaluate EPIC on mathematical reasoning, where it outperforms all baselines.

Autoregressive transformers are the backbone of modern large language models. Despite their success, inference remains slow due to strictly sequential prediction. Prior attempts to predict multiple tokens per step typically impose independence assumptions across tokens, which limits their ability to match the full expressiveness of standard autoregressive models. In this work, we break this paradigm by proposing an efficient and universal framework to jointly predict multiple tokens in a single transformer call, without limiting the representational power. Inspired by ideas from inverse autoregressive normalizing flows, we convert a series of random variables deterministically into a token sequence, incorporating the sampling procedure into a trained model. This allows us to train parallelized models both from scratch and by distilling an existing autoregressive model. Empirically, our distilled model matches its teacher's output for an average of close to 50 tokens on toy data and 5 tokens on a coding dataset, all within a single forward pass.

Agentic benchmarks increasingly rely on LLM-simulated users to scalably evaluate agent performance, yet the robustness, validity, and fairness of this approach remain unexamined. Through a user study with participants across the United States, India, Kenya, and Nigeria, we investigate whether LLM-simulated users serve as reliable proxies for real human users in evaluating agents on τ-Bench retail tasks. We find that user simulators lack robustness, with agent success rates varying up to 9 percentage points across different user LLMs. Furthermore, simulated users systematically miscalibrate performance, underestimating success on challenging tasks while overestimating moderately difficult ones. African American Vernacular English (AAVE) speakers experience consistently worse success rates and calibration errors than Standard American English (SAE) speakers, with disparities compounding significantly with age. We also find simulated users to be a differentially effective proxy for different populations, performing worst for AAVE and Indian English. Additionally, simulated users introduce conversational artifacts and surface different failure patterns than human users. These findings demonstrate that current evaluation practices risk misrepresenting agent capabilities across diverse user populations and may obscure real-world deployment challenges.

Humans increasingly turn to Language Models (LMs) in ways that shape beliefs and drive decisions, including discussing, rewriting, and summarizing information from scientific articles, news, and medical reports. However, in these domains, where how confidently a claim is expressed matters, little is known about whether LMs faithfully preserve it. In this work, we investigate certainty distortion in LMs, defined as meaningful changes in expressed certainty when semantic content is preserved. We propose an LM-based evaluation metric that is consistent with population-level judgments of certainty. Using this metric, we characterize certainty distortion across different sizes and families of models in the context of scientific and medical communication tasks. Our results show that certainty distortion affects up to 75% of LM outputs and is systematically asymmetric in rewriting tasks with most LMs being 1.5-2× more likely to increase the expressed certainty than to decrease it. These effects can compound over repeated paraphrasing: in the medical domain, claude-haiku-4-5 increases certainty of 20% examples after a single iteration, increasing to 40% after five iterations. Prompt-based interventions reduce overall certainty distortion but do not eliminate it. Together, these findings reveal a general bias toward inflating expressed certainty, with direct implications for users who rely on LMs in high-stakes domains.

Semantic control entails steering LM generations towards satisfying subtle non-lexical constraints, e.g., toxicity, sentiment, or politeness, attributes that can be captured by a sequence-level verifier. It can thus be viewed as sampling from the LM distribution conditioned on the target attribute, a computationally intractable problem due to the non-decomposable nature of the verifier. Existing approaches to LM control either only deal with syntactic constraints which cannot capture the aforementioned attributes, or rely on sampling to explore the conditional LM distribution, an ineffective estimator for low-probability events. In this work, we leverage a verifier's gradient information to efficiently reason over all generations that satisfy the target attribute, enabling precise steering of LM generations by reweighing the next-token distribution. Starting from an initial sample, we create a local LM distribution favoring semantically similar sentences. This approximation enables the tractable computation of an expected sentence embedding. We use this expected embedding, informed by the verifier's evaluation at the initial sample, to estimate the probability of satisfying the constraint, which directly informs the update to the next-token distribution. We evaluated the effectiveness of our approach in controlling the toxicity, sentiment, and topic-adherence of LMs yielding generations satisfying the constraint with high probability (>95

We propose Parallel Token Prediction (PTP), a universal framework for parallel sequence generation in language models. PTP jointly predicts multiple dependent tokens in a single transformer call by incorporating the sampling procedure into the model. This reduces the latency bottleneck of autoregressive decoding, and avoids the restrictive independence assumptions common in existing multi-token prediction methods. We prove that PTP can represent arbitrary autoregressive sequence distributions. PTP is trained either by distilling an existing model or through inverse autoregressive training without a teacher. Experimentally, we achieve state-of-the-art speculative decoding performance on Vicuna-7B by accepting over four tokens per step on Spec-Bench. The universality of our framework indicates that parallel generation of long sequences is feasible without loss of modeling power.

State space models (SSMs) are a promising alternative to transformers for language modeling because they use fixed memory during inference. However, this fixed memory usage requires some information loss in the hidden state when processing long sequences. While prior work has studied the sequence length at which this information loss occurs, it does not characterize the types of information SSM language models (LMs) tend to forget. In this paper, we address this knowledge gap by identifying the types of tokens (e.g., parts of speech, named entities) and sequences (e.g., code, math problems) that are more frequently forgotten by SSM LMs. We achieve this by training an auto-encoder to reconstruct sequences from the SSM's hidden state, and measure information loss by comparing inputs with their reconstructions. We perform experiments using the Mamba family of SSM LMs (130M–1.4B) on sequences ranging from 4–256 tokens. Our results show significantly higher rates of information loss on math-related tokens (e.g., numbers, variables), mentions of organization entities, and alternative dialects to Standard American English. We then examine the frequency that these tokens appear in Mamba's pretraining data and find that less prevalent tokens tend to be the ones Mamba is most likely to forget. By identifying these patterns, our work provides clear direction for future research to develop methods that better control Mamba's ability to retain important information.

Humans have the ability to reason about geometric patterns in images and scenes from a young age. However, developing large multimodal models (LMMs) capable of similar reasoning remains a challenge, highlighting the need for robust evaluation methods to assess these capabilities. We introduce TurtleBench, a benchmark designed to evaluate LMMs' capacity to interpret geometric patterns-given visual examples, textual instructions, or both- and generate precise code outputs. Inspired by turtle geometry, a notion used to teach children foundational coding and geometric concepts, TurtleBench features tasks with patterned shapes that have underlying algorithmic logic. Our evaluation reveals that leading LMMs struggle significantly with these tasks, with GPT-4o achieving only 19% accuracy on the simplest tasks and few-shot prompting only marginally improves their performance (< 2%). TurtleBench highlights the gap between human and AI performance in intuitive and visual geometrical understanding, setting the stage for future research in this area. TurtleBench stands as one of the few benchmarks to evaluate the integration of visual understanding and code generation capabilities in LMMs, setting the stage for future research. Code and Dataset for this paper is provided here: https://github.com/sinaris76/TurtleBench

Large language models (LLMs) require alignment to effectively and safely follow user instructions. This process necessitates training an aligned version for every base model, resulting in significant computational overhead. In this work, we propose NUDGING, a simple, training-free algorithm that aligns any base model at inference time using a small aligned model. NUDGING is motivated by recent findings that alignment primarily alters the model’s behavior on a small subset of stylistic tokens (e.g., discourse markers). We find that base models are significantly more uncertain when generating these tokens. Building on this insight, NUDGING employs a small aligned model to generate nudging tokens to guide the base model’s output during decoding when the base model’s uncertainty is high, with only a minor additional inference overhead. We evaluate NUDGING across 3 model families on a diverse range of open-instruction tasks. Without any training, nudging a large base model with a 7×-14× smaller aligned model achieves zero-shot performance comparable to, and sometimes surpassing, that of large aligned models. By operating at the token level, NUDGING enables off-the-shelf collaboration between model families. For instance, nudging Gemma-2-27b with Llama-27b-chat outperforms Llama-2-70b-chat on various tasks. Overall, our work offers a modular and cost-efficient solution to LLM alignment. Our code and demo are available at: https://fywalter.github.io/nudging/.

In-context learning (ICL) with dynamically selected demonstrations combines the flexibility of prompting large language models (LLMs) with the ability to leverage training data to improve performance. While ICL has been highly successful for prediction and generation tasks, leveraging it for agentic tasks that require sequential decision making is challenging – one must think not only about how to annotate long trajectories at scale and how to select demonstrations, but also what constitutes demonstrations, and when and where to show them. To address this, we first propose an algorithm that leverages an LLM with retries along with demonstrations to automatically and efficiently annotate agentic tasks with solution trajectories. We then show that set-selection of trajectories of similar tasks as demonstrations significantly improves performance, reliability, robustness, and efficiency of LLM agents. However, trajectory demonstrations have a large inference cost overhead. We show that this can be mitigated by using small trajectory snippets at every step instead of an additional trajectory. We find that demonstrations obtained from larger models (in the annotation phase) also improve smaller models, and that ICL agents can even rival costlier trained agents. Thus, our results reveal that ICL, with careful use, can be very powerful for agentic tasks as well.

The inevitable appearance of spurious correlations in training datasets hurts the generalization of NLP models on unseen data. Previous work has found that datasets with paired inputs are prone to correlations between a specific part of the input (e.g., the hypothesis in NLI) and the label; consequently, models trained only on those outperform chance. Are these correlations picked up by models trained on the full input data? To address this question, we propose a new evaluation method, Counterfactual Attentiveness Test (CAT). CAT uses counterfactuals by replacing part of the input with its counterpart from a different example (subject to some restrictions), expecting an attentive model to change its prediction. Using CAT, we systematically investigate established supervised and in-context learning models on ten datasets spanning four tasks: natural language inference, reading comprehension, paraphrase detection, and visual & language reasoning. CAT reveals that reliance on such correlations is mainly data-dependent. Surprisingly, we find that GPT3 becomes less attentive with an increased number of demonstrations, while its accuracy on the test data improves. Our results demonstrate that augmenting training or demonstration data with counterfactuals is effective in improving models' attentiveness. We show that models' attentiveness measured by CAT reveals different conclusions from solely measuring correlations in data.

Large language models (LLMs) have recently been used for sequential decision making in interactive environments. However, leveraging environment reward signals for continual LLM actor improvement is not straightforward. We propose Skill Set Optimization (SSO) for improving LLM actor performance through constructing and refining sets of transferable skills. SSO constructs skills by extracting common subtrajectories with high rewards and generating subgoals and instructions to represent each skill. These skills are provided to the LLM actor in-context to reinforce behaviors with high rewards. Then, SSO further refines the skill set by pruning skills that do not continue to result in high rewards. We evaluate our method in the classic videogame NetHack and the text environment ScienceWorld to demonstrate SSO's ability to optimize a set of skills and perform in-context policy improvement. SSO outperforms baselines by 40% in our custom NetHack task and outperforms the previous state-of-the-art in ScienceWorld by 35%.

Bias amplification is a phenomenon in which models exacerbate biases or stereotypes present in the training data. In this paper, we study bias amplification in the text-to-image domain using Stable Diffusion by comparing gender ratios in training vs. generated images. We find that the model appears to amplify gender-occupation biases found in the training data (LAION) considerably. However, we discover that amplification can be largely attributed to discrepancies between training captions and model prompts. For example, an inherent difference is that captions from the training data often contain explicit gender information while our prompts do not, which leads to a distribution shift and consequently inflates bias measures. Once we account for distributional differences between texts used for training and generation when evaluating amplification, we observe that amplification decreases drastically. Our findings illustrate the challenges of comparing biases in models and their training data, and highlight confounding factors that impact analyses.

*Uncertainty expressions* such as ‘probably’ or ‘highly unlikely’ are pervasive in human language. While prior work has established that there is population-level agreement in terms of how humans quantitatively interpret these expressions, there has been little inquiry into the abilities of language models in the same context. In this paper, we investigate how language models map linguistic expressions of uncertainty to numerical responses. Our approach assesses whether language models can employ theory of mind in this setting: understanding the uncertainty of another agent about a particular statement, independently of the model’s own certainty about that statement. We find that 7 out of 10 models are able to map uncertainty expressions to probabilistic responses in a human-like manner. However, we observe systematically different behavior depending on whether a statement is actually true or false. This sensitivity indicates that language models are substantially more susceptible to bias based on their prior knowledge (as compared to humans). These findings raise important questions and have broad implications for human-AI and AI-AI communication.

Misgendering, the act of incorrectly addressing someone's gender, inflicts serious harm and is pervasive in everyday technologies, yet there is a notable lack of research to combat it. We are the first to address this lack of research into interventions for misgendering by conducting a survey of gender-diverse individuals in the US to understand perspectives about automated interventions for text-based misgendering. Based on survey insights on the prevalence of misgendering, desired solutions, and associated concerns, we introduce a misgendering interventions task and evaluation dataset, MISGENDERMENDER. We define the task with two sub-tasks: (i) detecting misgendering, followed by (ii) correcting misgendering where misgendering is present in domains where editing is appropriate. MISGENDERMENDER comprises 3790 instances of social media content and LLM-generations about non-cisgender public figures, annotated for the presence of misgendering, with additional annotations for correcting misgendering in LLM-generated text. Using this dataset, we set initial benchmarks by evaluating existing NLP systems and highlighting challenges for future models to address. We release the full dataset, code, and demo at https://tamannahossainkay.github.io/misgendermender/.

Language models are achieving impressive performance on various tasks by aggressively adopting inference-time prompting techniques, such as zero-shot and few-shot prompting. In this work, we introduce EchoPrompt, a simple yet effective approach that prompts the model to rephrase its queries before answering them. EchoPrompt is adapted for both zero-shot and few-shot in-context learning with standard and chain-of-thought prompting. Experimental results show that EchoPrompt yields substantial improvements across all these settings for four families of causal language models. These improvements are observed across various numerical reasoning (e.g. GSM8K, SVAMP), reading comprehension (e.g. DROP), and logical reasoning (e.g. Coin Flipping) tasks. On average, EchoPrompt improves the Zero-shot-CoT performance of code-davinci-002 by 5% in numerical tasks and 13% in reading comprehension tasks. We investigate the factors contributing to EchoPrompt's effectiveness through ablation studies, which reveal that both the original query and the model-generated rephrased version are instrumental in its performance gains. Our empirical results indicate that EchoPrompt is an effective technique that enhances in-context learning performance. We recommend incorporating EchoPrompt into various baseline prompting strategies to achieve performance boosts.

Watermarking involves implanting an imperceptible signal into generated text that can later be detected via statistical tests. A prominent family of watermarking strategies for LLMs embeds this signal by upsampling a (pseudorandomly-chosen) subset of tokens at every generation step. However, such signals alter the model’s output distribution and can have unintended effects on its downstream performance. In this work, we evaluate the performance of LLMs watermarked using three different strategies over a diverse suite of tasks including those cast as k-class classification (CLS), multiple choice question answering (MCQ), short-form generation (e.g., open-ended question answering) and long-form generation (e.g., translation) tasks. We find that watermarks (under realistic hyperparameters) can cause significant drops in LLMs’ effective utility across all tasks. We observe drops of 10 to 20% in CLS tasks in the average case, which shoot up to 100% in the worst case. We notice degradations of about 7% in MCQ tasks, 10-15% in short-form generation, and 5-15% in long-form generation tasks. Our findings highlight the trade-offs that users should be cognizant of when using watermarked models.

Large text corpora are the backbone of language models. However, we have a limited understanding of the content of these corpora, including general statistics, quality, social factors, and inclusion of evaluation data (contamination). In this work, we propose What's In My Big Data? (WIMBD), a platform and a set of sixteen analyses that allow us to reveal and compare the contents of large text corpora. WIMBD builds on two basic capabilities -- count and search -- at scale, which allows us to analyze more than 35 terabytes on a standard compute node. We apply WIMBD to ten different corpora used to train popular language models, including C4, The Pile, and RedPajama. Our analysis uncovers several surprising and previously undocumented findings about these corpora, including the high prevalence of duplicate, synthetic, and low-quality content, personally identifiable information, toxic language, and benchmark contamination. For instance, we find that about 50% of the documents in RedPajama and LAION-2B-en are duplicates. In addition, several datasets used for benchmarking models trained on such corpora are contaminated with respect to important benchmarks, including the Winograd Schema Challenge and parts of GLUE and SuperGLUE. We open-source WIMBD's code and artifacts to provide a standard set of evaluations for new text-based corpora and to encourage more analyses and transparency around them.

Recent advances in multimodal models show remarkable performance in real-world benchmarks for chart and figure understanding like ChartQA that involve interpreting trends, comparing data points, and extracting insights from visuals.In this paper, we investigate the extent to which these models truly comprehend the underlying information in charts by posing direct, elementary questions about simple features such as axes ranges and values to examine their fundamental visual understanding abilities in the context of charts.Our questions are applied to two sets of figures: synthetic and real-world.The empirical evaluation of 5 popular multimodal models on our dataset reveals shortfalls in understanding charts and figures, contrary to what their performance on complex benchmarks might suggest.For instance, Gemini Pro Vision only achieves 57.9% accuracy on our elementary set of questions on real-world plots, while other popular multimodal models showed similar or less performance.This work highlights an important limitation of current multimodal models, and cautions against overly optimistic interpretations of their abilities based on results of canonical evaluations.