Language model outputs are not always reliable; this prompts research into methods for adapting model responses based on uncertainty. Common approaches include: \emph{abstention}, where models refrain from generating responses when uncertain; and \emph{linguistic calibration}, where models hedge their statements using uncertainty quantifiers. However, abstention can withhold valuable information, while linguistically calibrated responses are often challenging to leverage in downstream tasks. We propose a unifying view of both approaches, Conformal Linguistic Calibration (CLC), reinterpreting linguistic calibration as answer set prediction. We begin by presenting a unified framework that connects abstention and linguistic calibration through the lens of linguistic pragmatics. We then describe an implementation that allows for controlling the level of imprecision in model responses. Experimental results show that our method produces calibrated outputs with conformal guarantees on factual accuracy. Furthermore, our approach enables fine-tuning models to perform uncertainty-aware adaptive claim rewriting, offering a controllable balance between factuality and specificity.

Large language models (LLMs) can prove mathematical theorems formally by generating proof steps (\textit{a.k.a.} tactics) within a proof system. However, the space of possible tactics is vast and complex, while the available training data for formal proofs is limited, posing a significant challenge to LLM-based tactic generation. To address this, we introduce a neuro-symbolic tactic generator that synergizes the mathematical intuition learned by LLMs with domain-specific insights encoded by symbolic methods. The key aspect of this integration is identifying which parts of mathematical reasoning are best suited to LLMs and which to symbolic methods. While the high-level idea of neuro-symbolic integration is broadly applicable to various mathematical problems, in this paper, we focus specifically on Olympiad inequalities (Figure~1). We analyze how humans solve these problems and distill the techniques into two types of tactics: (1) scaling, handled by symbolic methods, and (2) rewriting, handled by LLMs. In addition, we combine symbolic tools with LLMs to prune and rank the proof goals for efficient proof search. We evaluate our framework on 161 challenging inequalities from multiple mathematics competitions, achieving state-of-the-art performance and significantly outperforming existing LLM and symbolic approaches without requiring additional training data.

The study explores whether Large Language Models (LLMs) exhibit negation-induced forgetting (NIF), a cognitive phenomenon observed in humans where negating incorrect attributes of an object or event leads to diminished recall of this object or event compared to affirming correct attributes (Mayo et al., 2014; Zang et al., 2023). We adapted Zang et al. (2023) experimental framework to test this effect in ChatGPT-3.5, GPT-4o mini and Llama3-70b-instruct. Our results show that ChatGPT-3.5 exhibits NIF, with negated information being less likely to be recalled than affirmed information. GPT-4o-mini showed a marginally significant NIF effect, while LLaMA-3-70B did not exhibit NIF. The findings provide initial evidence of negation-induced forgetting in some LLMs, suggesting that similar cognitive biases may emerge in these models. This work is a preliminary step in understanding how memory-related phenomena manifest in LLMs.

Deep Reinforcement Learning (RL) models often fail to generalize when even small changes occur in the environment's observations or task requirements. Addressing these shifts typically requires costly retraining, limiting the reusability of learned policies. In this paper, we build on recent work in semantic alignment to propose a zero-shot method for mapping between latent spaces across different agents trained on different visual and task variations. Specifically, we learn a transformation that maps embeddings from one agent's encoder to another agent's encoder without further fine-tuning. Our approach relies on a small set of "anchor" observations that are semantically aligned, which we use to estimate an affine or orthogonal transform. Once the transformation is found, an existing controller trained for one domain can interpret embeddings from a different (existing) encoder in a zero-shot fashion, skipping additional trainings. We empirically demonstrate that our framework preserves high performance under visual and task domain shifts. We empirically demonstrate zero-shot stitching performance on the CarRacing environment with changing background and task. By allowing modular re-assembly of existing policies, it paves the way for more robust, compositional RL in dynamically changing environments.

Evaluating the pedagogical capabilities of AI-based tutoring models is critical for making guided progress in the field. Yet, we lack a reliable, easy-to-use, and simple-to-run evaluation that reflects the pedagogical abilities of models. To fill this gap, we present MathTutorBench, an open-source benchmark for holistic tutoring model evaluation. MathTutorBench contains a collection of datasets and metrics that broadly cover tutor abilities as defined by learning sciences research in dialog-based teaching. To score the pedagogical quality of open-ended teacher responses, we train a reward model and show it can discriminate expert from novice teacher responses with high accuracy. We evaluate a wide set of closed- and open-weight models on MathTutorBench and find that subject expertise, indicated by solving ability, does not immediately translate to good teaching. Rather, pedagogy and subject expertise appear to form a trade-off that is navigated by the degree of tutoring specialization of the model. Furthermore, tutoring appears to become more challenging in longer dialogs, where simpler questioning strategies begin to fail. We release the benchmark, code, and leaderboard openly to enable rapid benchmarking of future models.

Large Language Models (LLMs) offer natural language explanations as an alternative to feature attribution methods for model interpretability. However, despite their plausibility, they may not reflect the model's internal reasoning faithfully, which is crucial for understanding the model's true decision-making processes. Although several faithfulness metrics have been proposed, a unified evaluation framework remains absent. To address this gap, we present Causal Diagnosticity, a framework to evaluate faithfulness metrics for natural language explanations. Our framework employs the concept of causal diagnosticity, and uses model-editing methods to generate faithful-unfaithful explanation pairs. Our benchmark includes four tasks: fact-checking, analogy, object counting, and multi-hop reasoning. We evaluate a variety of faithfulness metrics, including post-hoc explanation and chain-of-thought-based methods. We find that all tested faithfulness metrics often fail to surpass a random baseline. Our work underscores the need for improved metrics and more reliable interpretability methods in LLMs.

Providing effective feedback for programming assignments in computer science education can be challenging: students solve problems by iteratively submitting code, executing it, and using limited feedback from the compiler or the auto-grader to debug. Analyzing student debugging behavior in this process may reveal important insights into their knowledge and inform better personalized support tools. In this work, we propose an encoder-decoder-based model that learns meaningful code-edit embeddings between consecutive student code submissions, to capture their debugging behavior. Our model leverages information on whether a student code submission passes each test case to fine-tune large language models (LLMs) to learn code editing representations. It enables personalized next-step code suggestions that maintain the student's coding style while improving test case correctness. Our model also enables us to analyze student code-editing patterns to uncover common student errors and debugging behaviors, using clustering techniques. Experimental results on a real-world student code submission dataset demonstrate that our model excels at code reconstruction and personalized code suggestion while revealing interesting patterns in student debugging behavior.

Multimodal Large Language Models (MLLMs), built upon LLMs, have recently gained attention for their capabilities in image recognition and understanding. However, while MLLMs are vulnerable to adversarial attacks, the transferability of these attacks across different models remains limited, especially under targeted attack setting. Existing methods primarily focus on vision-specific perturbations but struggle with the complex nature of vision-language modality alignment. In this work, we introduce the Dynamic Vision-Language Alignment (DynVLA) Attack, a novel approach that injects dynamic perturbations into the vision-language connector to enhance generalization across diverse vision-language alignment of different models. Our experimental results show that DynVLA significantly improves the transferability of adversarial examples across various MLLMs, including BLIP2, InstructBLIP, MiniGPT4, LLaVA, and closed-source models such as Gemini.

Training dense LLMs requires enormous amounts of data and centralized compute, which introduces fundamental bottlenecks and ever-growing costs for large models. Several studies aim to reduce this dependency on centralization by reducing the communication overhead of training dense models. Taking this idea of reducing communication overhead to a natural extreme, by training embarrassingly parallelizable ensembles of small independent experts, has been shown to outperform large dense models trained in traditional centralized settings. However, existing studies do not take into account underlying differences amongst data domains and treat them as monolithic, regardless of their underlying complexity, size, or distribution. In this paper, we explore the effects of introducing heterogeneity to these ensembles of domain expert models. Specifically, by allowing models within the ensemble to vary in size--as well as the number of training steps taken depending on the training data's domain--we study the effect heterogeneity has on these ensembles when evaluated against domains included in, and excluded from, the training set. We use the same compute budget to train heterogeneous ensembles and homogeneous baselines for comparison. We show that the heterogeneous ensembles achieve the lowest perplexity scores in $20$ out of the $21$ data domains used in the evaluation. Our code is available at https://github.com/gensyn-ai/hdee.

Modern foundation models rely heavily on using scaling laws to guide crucial training decisions. Researchers often extrapolate the optimal architecture and hyper parameters settings from smaller training runs by describing the relationship between, loss, or task performance, and scale. All components of this process vary, from the specific equation being fit, to the training setup, to the optimization method. Each of these factors may affect the fitted law, and therefore, the conclusions of a given study. We discuss discrepancies in the conclusions that several prior works reach, on questions such as the optimal token to parameter ratio. We augment this discussion with our own analysis of the critical impact that changes in specific details may effect in a scaling study, and the resulting altered conclusions. Additionally, we survey over 50 papers that study scaling trends: while 45 of these papers quantify these trends using a power law, most under-report crucial details needed to reproduce their findings. To mitigate this, we we propose a checklist for authors to consider while contributing to scaling law research.