Despite significant progress, recent studies indicate that current large language models (LLMs) may still capture dataset biases and utilize them during inference, leading to the poor generalizability of LLMs. However, due to the diversity of dataset biases and the insufficient nature of bias suppression based on in-context learning, the effectiveness of previous prior knowledge-based debiasing methods and in-context learning based automatic debiasing methods is limited. To address these challenges, we explore the combination of causal mechanisms with information theory and propose an information gain-guided causal intervention debiasing (IGCIDB) framework. This framework first utilizes an information gain-guided causal intervention method to automatically and autonomously balance the distribution of instruction-tuning dataset. Subsequently, it employs a standard supervised fine-tuning process to train LLMs on the debiased dataset. Experimental results show that IGCIDB can effectively debias LLM to improve its generalizability across different tasks.
Pretraining data has a direct impact on the behaviors and quality of language models (LMs), but we only understand the most basic principles of this relationship. While most work focuses on pretraining data's effect on downstream task behavior, we investigate its relationship to LM representations. Previous work has discovered that, in language models, some concepts are encoded `linearly' in the representations, but what factors cause these representations to form? We study the connection between pretraining data frequency and models' linear representations of factual relations. We find evidence that the formation of linear representations is strongly connected to pretraining term frequencies; specifically for subject-relation-object fact triplets, both subject-object co-occurrence frequency and in-context learning accuracy for the relation are highly correlated with linear representations. This is the case across all phases of pretraining. In OLMo-7B and GPT-J, we discover that a linear representation consistently (but not exclusively) forms when the subjects and objects within a relation co-occur at least 1k and 2k times, respectively, regardless of when these occurrences happen during pretraining. Finally, we train a regression model on measurements of linear representation quality in fully-trained LMs that can predict how often a term was seen in pretraining. Our model achieves low error even on inputs from a different model with a different pretraining dataset, providing a new method for estimating properties of the otherwise-unknown training data of closed-data models. We conclude that the strength of linear representations in LMs contains signal about the models' pretraining corpora that may provide new avenues for controlling and improving model behavior: particularly, manipulating the models' training data to meet specific frequency thresholds.
Transformer models exhibit remarkable in-context learning (ICL), adapting to novel tasks from examples within their context, yet the underlying mechanisms remain largely mysterious. Here, we provide an exact analytical characterization of ICL emergence by deriving the closed-form stochastic gradient descent (SGD) dynamics for a simplified linear transformer performing regression tasks. Our analysis reveals key properties: (1) a natural separation of timescales directly governed by the input data's covariance structure, leading to staged learning; (2) an exact description of how ICL develops, including fixed points corresponding to learned algorithms and conservation laws constraining the dynamics; and (3) surprisingly nonlinear learning behavior despite the model's linearity. We hypothesize this phenomenology extends to non-linear models. To test this, we introduce theory-inspired macroscopic measures (spectral rank dynamics, subspace stability) and use them to provide mechanistic explanations for (1) the sudden emergence of ICL in attention-only networks and (2) delayed generalization (grokking) in modular arithmetic models. Our work offers an exact dynamical model for ICL and theoretically grounded tools for analyzing complex transformer training.
Long-context capabilities are essential for a wide range of applications, including document and video understanding, in-context learning, and inference-time scaling, all of which require models to process and reason over long sequences of text and multimodal data. In this work, we introduce a efficient training recipe for building ultra-long context LLMs from aligned instruct model, pushing the boundaries of context lengths from 128K to 1M, 2M, and 4M tokens. Our approach leverages efficient continued pretraining strategies to extend the context window and employs effective instruction tuning to maintain the instruction-following and reasoning abilities. Our UltraLong-8B, built on Llama3.1-Instruct with our recipe, achieves state-of-the-art performance across a diverse set of long-context benchmarks. Importantly, models trained with our approach maintain competitive performance on standard benchmarks, demonstrating balanced improvements for both long and short context tasks. We further provide an in-depth analysis of key design choices, highlighting the impacts of scaling strategies and data composition. Our findings establish a robust framework for efficiently scaling context lengths while preserving general model capabilities. We release all model weights at: https://ultralong.github.io/.
Conditional layout generation aims to automatically generate visually appealing and semantically coherent layouts from user-defined constraints. While recent methods based on generative models have shown promising results, they typically require substantial amounts of training data or extensive fine-tuning, limiting their versatility and practical applicability. Alternatively, some training-free approaches leveraging in-context learning with Large Language Models (LLMs) have emerged, but they often suffer from limited reasoning capabilities and overly simplistic ranking mechanisms, which restrict their ability to generate consistently high-quality layouts. To this end, we propose LayoutCoT, a novel approach that leverages the reasoning capabilities of LLMs through a combination of Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) techniques. Specifically, LayoutCoT transforms layout representations into a standardized serialized format suitable for processing by LLMs. A Layout-aware RAG is used to facilitate effective retrieval and generate a coarse layout by LLMs. This preliminary layout, together with the selected exemplars, is then fed into a specially designed CoT reasoning module for iterative refinement, significantly enhancing both semantic coherence and visual quality. We conduct extensive experiments on five public datasets spanning three conditional layout generation tasks. Experimental results demonstrate that LayoutCoT achieves state-of-the-art performance without requiring training or fine-tuning. Notably, our CoT reasoning module enables standard LLMs, even those without explicit deep reasoning abilities, to outperform specialized deep-reasoning models such as deepseek-R1, highlighting the potential of our approach in unleashing the deep reasoning capabilities of LLMs for layout generation tasks.
Language models rely on semantic priors to perform in-context learning, which leads to poor performance on tasks involving inductive reasoning. Instruction-tuning methods based on imitation learning can superficially enhance the in-context learning performance of language models, but they often fail to improve the model's understanding of the underlying rules that connect inputs and outputs in few-shot demonstrations. We propose ReDis, a reasoning distillation technique designed to improve the inductive reasoning capabilities of language models. Through a careful combination of data augmentation, filtering, supervised fine-tuning, and alignment, ReDis achieves significant performance improvements across a diverse range of tasks, including 1D-ARC, List Function, ACRE, and MiniSCAN. Experiments on three language model backbones show that ReDis outperforms equivalent few-shot prompting baselines across all tasks and even surpasses the teacher model, GPT-4o, in some cases. ReDis, based on the LLaMA-3 backbone, achieves relative improvements of 23.2%, 2.8%, and 66.6% over GPT-4o on 1D-ARC, ACRE, and MiniSCAN, respectively, within a similar hypothesis search space. The code, dataset, and model checkpoints will be made available at https://github.com/NafisSadeq/reasoning-distillation.git.
Automated analysis for engineering structures offers considerable potential for boosting efficiency by minimizing repetitive tasks. Although AI-driven methods are increasingly common, no systematic framework yet leverages Large Language Models (LLMs) for automatic structural analysis. To address this gap, we propose a novel framework that integrates LLMs with structural analysis software. LLMs serve as the core engine: they parse structural descriptions from text and translate them into executable Python scripts. Moreover, the framework integrates the generative capabilities of LLMs with code-based finite element (FE) tools like OpenSeesPy. It employs domain-specific prompt design and in-context learning strategies to enhance the LLM's problem-solving capabilities and generative stability, enabling fully automated structural analysis from descriptive text to model outputs. In our experiments, we introduce a well-curated small-scale benchmark dataset of 20 structural analysis word problems (SAWPs) with ground-truth solutions and evaluate the performance of different LLMs within our framework in solving these SAWPs. The role of system instructions, crafted by structural engineers, is also investigated to understand their impact on LLM-driven structural analysis. Additionally, the generative stability of our framework is examined. Through multiple validation experiments on the benchmark, our results demonstrate that the proposed framework can substantially increase the level of automation in solving SAWPs compared to traditional methods. Quantitatively, the framework, built on GPT-4o, achieved 100% accuracy, surpassing GPT-4 (85%), Gemini 1.5 Pro (80%), and Llama-3.3 (30%) on the test examples. Furthermore, integrating domain-specific instructions enhanced performance by 30% on problems with asymmetrical structural configurations.
Detecting empathy from video interactions is an emerging area of research. Video datasets, however, are often released as extracted features (i.e., tabular data) rather than raw footage due to privacy and ethical concerns. Prior research on such tabular datasets established tree-based classical machine learning approaches as the best-performing models. Motivated by the recent success of textual foundation models (i.e., large language models), we explore the use of tabular foundation models in empathy detection from tabular visual features. We experiment with two recent tabular foundation models $-$ TabPFN v2 and TabICL $-$ through in-context learning and fine-tuning setups. Our experiments on a public human-robot interaction benchmark demonstrate a significant boost in cross-subject empathy detection accuracy over several strong baselines (accuracy: $0.590 \rightarrow 0.730$; AUC: $0.564 \rightarrow 0.669$). In addition to performance improvement, we contribute novel insights and an evaluation setup to ensure generalisation on unseen subjects in this public benchmark. As the practice of releasing video features as tabular datasets is likely to persist due to privacy constraints, our findings will be widely applicable to future empathy detection video datasets as well.
Large Reasoning Models (LRMs) like DeepSeek-R1 and OpenAI-o1 have demonstrated remarkable reasoning capabilities, raising important questions about their biases in LLM-as-a-judge settings. We present a comprehensive benchmark comparing judging biases between LLMs and LRMs across both subjective preference-alignment datasets and objective fact-based datasets. Through investigation of bandwagon, authority, position, and distraction biases, we uncover four key findings: (1) despite their advanced reasoning capabilities, LRMs remain susceptible to the above biases; (2) LRMs demonstrate better robustness than LLMs specifically on fact-related datasets; (3) LRMs exhibit notable position bias, preferring options in later positions; and (4) we identify a novel "superficial reflection bias" where phrases mimicking reasoning (e.g., "wait, let me think...") significantly influence model judgments. To address these biases, we design and evaluate three mitigation strategies: specialized system prompts that reduce judging biases by up to 19\% in preference alignment datasets and 14\% in fact-related datasets, in-context learning that provides up to 27\% improvement on preference tasks but shows inconsistent results on factual tasks, and a self-reflection mechanism that reduces biases by up to 10\% in preference datasets and 16\% in fact-related datasets, with self-reflection proving particularly effective for LRMs. Our work provides crucial insights for developing more reliable LLM-as-a-Judge frameworks, especially as LRMs become increasingly deployed as automated judges.
Many-shot jailbreaking (MSJ) is an adversarial technique that exploits the long context windows of modern LLMs to circumvent model safety training by including in the prompt many examples of a "fake" assistant responding inappropriately before the final request. With enough examples, the model's in-context learning abilities override its safety training, and it responds as if it were the "fake" assistant. In this work, we probe the effectiveness of different fine-tuning and input sanitization approaches on mitigating MSJ attacks, alone and in combination. We find incremental mitigation effectiveness for each, and show that the combined techniques significantly reduce the effectiveness of MSJ attacks, while retaining model performance in benign in-context learning and conversational tasks. We suggest that our approach could meaningfully ameliorate this vulnerability if incorporated into model safety post-training.
Heavy-duty trucks pose significant safety challenges due to their large size and limited maneuverability compared to passenger vehicles. A deeper understanding of truck characteristics is essential for enhancing the safety perspective of cooperative autonomous driving. Traditional LiDAR-based truck classification methods rely on extensive manual annotations, which makes them labor-intensive and costly. The rapid advancement of large language models (LLMs) trained on massive datasets presents an opportunity to leverage their few-shot learning capabilities for truck classification. However, existing vision-language models (VLMs) are primarily trained on image datasets, which makes it challenging to directly process point cloud data. This study introduces a novel framework that integrates roadside LiDAR point cloud data with VLMs to facilitate efficient and accurate truck classification, which supports cooperative and safe driving environments. This study introduces three key innovations: (1) leveraging real-world LiDAR datasets for model development, (2) designing a preprocessing pipeline to adapt point cloud data for VLM input, including point cloud registration for dense 3D rendering and mathematical morphological techniques to enhance feature representation, and (3) utilizing in-context learning with few-shot prompting to enable vehicle classification with minimally labeled training data. Experimental results demonstrate encouraging performance of this method and present its potential to reduce annotation efforts while improving classification accuracy.
Recently, there has been a growing trend in utilizing large language models (LLMs) for recommender systems, referred to as LLMRec. A notable approach within this trend is not to fine-tune these models directly but instead to leverage In-Context Learning (ICL) methods tailored for LLMRec, denoted as LLM-ICL Rec. Many contemporary techniques focus on harnessing ICL content to enhance LLMRec performance. However, optimizing LLMRec with ICL content presents unresolved challenges. Specifically, two key issues stand out: (1) the limited understanding of why using a few demonstrations without model fine-tuning can lead to better performance compared to zero-shot recommendations. (2) the lack of evaluation metrics for demonstrations in LLM-ICL Rec and the absence of the theoretical analysis and practical design for optimizing the generation of ICL content for recommendation contexts. To address these two main issues, we propose a theoretical model, the LLM-ICL Recommendation Equivalent Gradient Descent model (LRGD) in this paper, which connects recommendation generation with gradient descent dynamics. We demonstrate that the ICL inference process in LLM aligns with the training procedure of its dual model, producing token predictions equivalent to the dual model's testing outputs. Building on these theoretical insights, we propose an evaluation metric for assessing demonstration quality. We integrate perturbations and regularizations in LRGD to enhance the robustness of the recommender system. To further improve demonstration effectiveness, prevent performance collapse, and ensure long-term adaptability, we also propose a two-stage optimization process in practice. Extensive experiments and detailed analysis on three Amazon datasets validate the theoretical equivalence and support the effectiveness of our theoretical analysis and practical module design.
Mixture-of-Experts (MoE) Large Language Models (LLMs) suffer from severely sub-optimal expert pathways-our study reveals that naive expert selection learned from pretraining leaves a surprising 10-20% accuracy gap for improvement. Motivated by this observation, we develop a novel class of test-time optimization methods to re-weight or "re-mixing" the experts in different layers jointly for each test sample. Since the test sample's ground truth is unknown, we propose to optimize a surrogate objective defined by the sample's "successful neighbors" from a reference set of samples. We introduce three surrogates and algorithms based on mode-finding, kernel regression, and the average loss of similar reference samples/tasks. To reduce the cost of optimizing whole pathways, we apply our algorithms merely to the core experts' mixing weights in critical layers, which enjoy similar performance but save significant computation. This leads to "Critical-Layer, Core-Expert, Collaborative Pathway Optimization (C3PO)". We apply C3PO to two recent MoE LLMs and examine it on six widely-used benchmarks. It consistently improves the base model by 7-15% in accuracy and outperforms widely used test-time learning baselines, e.g., in-context learning and prompt/prefix tuning, by a large margin. Moreover, C3PO enables MoE LLMs with 1-3B active parameters to outperform LLMs of 7-9B parameters, hence improving MoE's advantages on efficiency. Our thorough ablation study further sheds novel insights on achieving test-time improvement on MoE.
Recent progress in diffusion models significantly advances various image generation tasks. However, the current mainstream approach remains focused on building task-specific models, which have limited efficiency when supporting a wide range of different needs. While universal models attempt to address this limitation, they face critical challenges, including generalizable task instruction, appropriate task distributions, and unified architectural design. To tackle these challenges, we propose VisualCloze, a universal image generation framework, which supports a wide range of in-domain tasks, generalization to unseen ones, unseen unification of multiple tasks, and reverse generation. Unlike existing methods that rely on language-based task instruction, leading to task ambiguity and weak generalization, we integrate visual in-context learning, allowing models to identify tasks from visual demonstrations. Meanwhile, the inherent sparsity of visual task distributions hampers the learning of transferable knowledge across tasks. To this end, we introduce Graph200K, a graph-structured dataset that establishes various interrelated tasks, enhancing task density and transferable knowledge. Furthermore, we uncover that our unified image generation formulation shared a consistent objective with image infilling, enabling us to leverage the strong generative priors of pre-trained infilling models without modifying the architectures.
Linear attention methods offer a compelling alternative to softmax attention due to their efficiency in recurrent decoding. Recent research has focused on enhancing standard linear attention by incorporating gating while retaining its computational benefits. Such Gated Linear Attention (GLA) architectures include competitive models such as Mamba and RWKV. In this work, we investigate the in-context learning capabilities of the GLA model and make the following contributions. We show that a multilayer GLA can implement a general class of Weighted Preconditioned Gradient Descent (WPGD) algorithms with data-dependent weights. These weights are induced by the gating mechanism and the input, enabling the model to control the contribution of individual tokens to prediction. To further understand the mechanics of this weighting, we introduce a novel data model with multitask prompts and characterize the optimization landscape of learning a WPGD algorithm. Under mild conditions, we establish the existence and uniqueness (up to scaling) of a global minimum, corresponding to a unique WPGD solution. Finally, we translate these findings to explore the optimization landscape of GLA and shed light on how gating facilitates context-aware learning and when it is provably better than vanilla linear attention.
The human face plays a central role in social communication, necessitating the use of performant computer vision tools for human-centered applications. We propose Face-LLaVA, a multimodal large language model for face-centered, in-context learning, including facial expression and attribute recognition. Additionally, Face-LLaVA is able to generate natural language descriptions that can be used for reasoning. Leveraging existing visual databases, we first developed FaceInstruct-1M, a face-centered database for instruction tuning MLLMs for face processing. We then developed a novel face-specific visual encoder powered by Face-Region Guided Cross-Attention that integrates face geometry with local visual features. We evaluated the proposed method across nine different datasets and five different face processing tasks, including facial expression recognition, action unit detection, facial attribute detection, age estimation and deepfake detection. Face-LLaVA achieves superior results compared to existing open-source MLLMs and competitive performance compared to commercial solutions. Our model output also receives a higher reasoning rating by GPT under a zero-shot setting across all the tasks. Both our dataset and model wil be released at https://face-llava.github.io to support future advancements in social AI and foundational vision-language research.
Error detection (ED) in tabular data is crucial yet challenging due to diverse error types and the need for contextual understanding. Traditional ED methods often rely heavily on manual criteria and labels, making them labor-intensive. Large language models (LLM) can minimize human effort but struggle with errors requiring a comprehensive understanding of data context. In this paper, we propose ZeroED, a novel hybrid zero-shot error detection framework, which combines LLM reasoning ability with the manual label-based ED pipeline. ZeroED operates in four steps, i.e., feature representation, error labeling, training data construction, and detector training. Initially, to enhance error distinction, ZeroED generates rich data representations using error reason-aware binary features, pre-trained embeddings, and statistical features. Then, ZeroED employs LLM to label errors holistically through in-context learning, guided by a two-step reasoning process for detailed error detection guidelines. To reduce token costs, LLMs are applied only to representative data selected via clustering-based sampling. High-quality training data is constructed through in-cluster label propagation and LLM augmentation with verification. Finally, a classifier is trained to detect all errors. Extensive experiments on seven public datasets demonstrate that, ZeroED substantially outperforms state-of-the-art methods by a maximum 30% improvement in F1 score and up to 90% token cost reduction.
Prior work on in-context copying has shown the existence of induction heads, which attend to and promote individual tokens during copying. In this work we introduce a new type of induction head: concept-level induction heads, which copy entire lexical units instead of individual tokens. Concept induction heads learn to attend to the ends of multi-token words throughout training, working in parallel with token-level induction heads to copy meaningful text. We show that these heads are responsible for semantic tasks like word-level translation, whereas token induction heads are vital for tasks that can only be done verbatim, like copying nonsense tokens. These two "routes" operate independently: in fact, we show that ablation of token induction heads causes models to paraphrase where they would otherwise copy verbatim. In light of these findings, we argue that although token induction heads are vital for specific tasks, concept induction heads may be more broadly relevant for in-context learning.
Program-guided reasoning has shown promise in complex claim fact-checking by decomposing claims into function calls and executing reasoning programs. However, prior work primarily relies on few-shot in-context learning (ICL) with ad-hoc demonstrations, which limit program diversity and require manual design with substantial domain knowledge. Fundamentally, the underlying principles of effective reasoning program generation still remain underexplored, making it challenging to construct effective demonstrations. To address this, we propose BOOST, a bootstrapping-based framework for few-shot reasoning program generation. BOOST explicitly integrates claim decomposition and information-gathering strategies as structural guidance for program generation, iteratively refining bootstrapped demonstrations in a strategy-driven and data-centric manner without human intervention. This enables a seamless transition from zero-shot to few-shot strategic program-guided learning, enhancing interpretability and effectiveness. Experimental results show that BOOST outperforms prior few-shot baselines in both zero-shot and few-shot settings for complex claim verification.
Large Vision-Language Models offer a new paradigm for AI-driven image understanding, enabling models to perform tasks without task-specific training. This flexibility holds particular promise across medicine, where expert-annotated data is scarce. Yet, VLMs' practical utility in intervention-focused domains--especially surgery, where decision-making is subjective and clinical scenarios are variable--remains uncertain. Here, we present a comprehensive analysis of 11 state-of-the-art VLMs across 17 key visual understanding tasks in surgical AI--from anatomy recognition to skill assessment--using 13 datasets spanning laparoscopic, robotic, and open procedures. In our experiments, VLMs demonstrate promising generalizability, at times outperforming supervised models when deployed outside their training setting. In-context learning, incorporating examples during testing, boosted performance up to three-fold, suggesting adaptability as a key strength. Still, tasks requiring spatial or temporal reasoning remained difficult. Beyond surgery, our findings offer insights into VLMs' potential for tackling complex and dynamic scenarios in clinical and broader real-world applications.