Despite growing interest in Theory of Mind (ToM) tasks for evaluating language models (LMs), little is known about how LMs internally represent mental states of self and others. Understanding these internal mechanisms is critical - not only to move beyond surface-level performance, but also for model alignment and safety, where subtle misattributions of mental states may go undetected in generated outputs. In this work, we present the first systematic investigation of belief representations in LMs by probing models across different scales, training regimens, and prompts - using control tasks to rule out confounds. Our experiments provide evidence that both model size and fine‐tuning substantially improve LMs’ internal representations of others’ beliefs, which are structured - not mere by-products of spurious correlations - yet brittle to prompt variations. Crucially, we show that these representations can be strengthened: targeted edits to model activations can correct wrong ToM inferences.
论文检索
输入标题、作者或关键词，从 960 篇学术成果中精准定位
Auto-regressive Large Language Models (LLMs) demonstrate remarkable performance across different domains such as vision and language tasks. However, due to sequential processing through multiple transformer layers, autoregressive decoding faces significant computational challenges, particularly in resource-constrained environments like mobile and edge devices. Existing approaches in literature that aim to improve latency via skipping layers have two distinct flavors: (1) early exit, and (2) input-agnostic heuristics where tokens exit at pre-determined layers irrespective of input sequence. Both the above strategies have limitations, the former cannot be applied in the presence of KV caching, which is essential for speed-ups in modern inference frameworks, and the latter fails to capture variation in layer importance across tasks or, more generally, across input sequences. To address these limitations, we propose FiRST, a model-agnostic framework that reduces inference latency by using layer-specific routers to adaptively skip transformer layers during decoding, based on routing decisions made from the input prompt in the prefill stage. FiRST remains fully compatible with KV caching, enabling faster decoding while maintaining quality. Our method reveals that input adaptivity is essential: Different tasks rely on different subsets of layers to evolve meaningful representations. Extensive experiments show that FiRST significantly reduces latency while outperforming existing layer selection strategies in quality. It retains performance comparable to the base model without skipping. FiRST is thus a promising and efficient solution for LLM deployment in low-resource environments.
We study how large language models (LLMs) reason about memorized knowledge through simple binary relations such as equality (=), inequality (<), and inclusion (\subset). Unlike in-context reasoning, the axioms (e.g., a < b, b < c) are only seen during training and not provided in the task prompt (e.g., evaluating a < c). The tasks require one or more reasoning steps, and data aggregation from one or more sources, showing performance change with task complexity. We introduce a lightweight technique, out-of-context representation learning, which trains only new token embeddings on axioms and evaluates them on unseen tasks. Across reflexivity, symmetry, and transitivity tests, LLMs mostly perform statistically significant better than chance, making the correct answer extractable when testing multiple phrasing variations, but still fall short of consistent reasoning on every single query. Analysis shows that the learned embeddings are organized in structured ways, suggesting real relational understanding. Surprisingly, it also indicates that the core reasoning happens during the training, not inference.
With the widespread application of Large Language Models (LLMs) in the field of Natural Language Processing (NLP), enhancing their performance has become a research hotspot. This paper presents a novel multi-prompt ensemble decoding approach designed to bolster the generation quality of LLMs by leveraging the aggregation of outcomes from multiple prompts. Given a unique input X, we submit n variations of prompts with X to LLMs in batch mode to decode and derive probability distributions. For each token prediction, we calculate the ensemble probability by averaging the n probability distributions within the batch, utilizing this aggregated probability to generate the token. This technique is dubbed Inner-Batch Ensemble. To facilitate efficient batch inference, we implement a Left-Padding strategy to maintain uniform input lengths across the n prompts. Through extensive experimentation on diverse NLP tasks, including code generation, text simplification and machine translation, we demonstrate the efficacy of our method in enhancing LLM performance. The results show substantial improvements in pass@k rates, LENS metrics and BLEU scores over conventional methods.
Event argument extraction identifies arguments for predefined event roles in text. Existing work evaluates this task with exact match (EM), where predicted arguments must align exactly with annotated spans. While suitable for span-based models, this approach falls short for large language models (LLMs), which often generate diverse yet semantically accurate arguments. EM severely underestimates performance by disregarding valid variations. Furthermore, EM evaluation fails to capture implicit arguments (unstated but inferable) and scattered arguments (distributed across a document). These limitations underscore the need for an evaluation framework that better captures models’ actual performance. To bridge this gap, we introduce REGen, a Reliable Evaluation framework for Generative event argument extraction. REGen combines the strengths of exact, relaxed, and LLM-based matching to better align with human judgment. Experiments on six datasets show that REGen reveals an average performance gain of +23.93 F1 over EM, reflecting capabilities overlooked by prior evaluation. Human validation further confirms REGen’s effectiveness, achieving 87.67% alignment with human assessments of argument correctness.
We propose a training-free approach to improve sentence embeddings leveraging test-time compute by applying generative text models for data augmentation at inference time. Unlike conventional data augmentation that utilises synthetic training data, our approach does not require access to model parameters or the computational resources typically required for fine-tuning state-of-the-art models. Generatively Augmented Sentence Encoding variates the input text by paraphrasing, summarising, or extracting keywords, followed by pooling the original and synthetic embeddings.Experimental results on the Massive Text Embedding Benchmark for Semantic Textual Similarity (STS) demonstrate performance improvements across a range of embedding models using different generative models for augmentation. We find that generative augmentation leads to larger performance improvements for embedding models with lower baseline performance. These findings suggest that integrating generative augmentation at inference time adds semantic diversity and can enhance the robustness and generalisability of sentence embeddings for embedding models. Our results show that performance gains depend on the embedding model and the dataset.
LiTEx: A Linguistic Taxonomy of Explanations for Understanding Within-Label Variation in Natural Language Inference
PDF ↗There is increasing evidence of Human Label Variation (HLV) in Natural Language Inference (NLI), where annotators assign different labels to the same premise-hypothesis pair. However, *within-label variation* — cases where annotators agree on the same label but provide divergent reasoning — poses an additional and mostly overlooked challenge. Several NLI datasets contain highlighted words in the NLI item as explanations, but the same spans on the NLI item can be highlighted for different reasons, as evidenced by free-text explanations, which offer a window into annotators’ reasoning. To systematically understand this problem and gain insight into the rationales behind NLI labels, we introduce LiTEx, a linguistically-informed taxonomy for categorizing free-text explanations in English. Using this taxonomy, we annotate a subset of the e-SNLI dataset, validate the taxonomy’s reliability, and analyze how it aligns with NLI labels, highlights, and explanations. We further assess the taxonomy’s usefulness in explanation generation, demonstrating that conditioning generation on LiTEx yields explanations that are linguistically closer to human explanations than those generated using only labels or highlights. Our approach thus not only captures within-label variation but also shows how taxonomy-guided generation for reasoning can bridge the gap between human and model explanations more effectively than existing strategies.
The growing deployment of deep learning models in real-world applications necessitates not only high predictive accuracy, but also mechanism to identify unreliable predictions, especially in high-stakes scenarios where decision risk must be minimized. Existing methods estimate uncertainty by leveraging predictive confidence (e.g., Softmax Response), structural characteristics of representation space (e.g., Mahalanobis distance), or stochastic variation in model outputs (e.g., Bayesian inference techniques such as Monte Carlo Dropout). In this work, we propose a novel uncertainty estimation (UE) framework based on sparse dictionary learning by identifying dictionary atoms associated with misclassified samples. We leverage pointwise mutual information (PMI) to quantify the association between sparse features and predictive failure. Our method – Sparsity-based Uncertainty Estimation (SUE) – is computationally efficient, offers interpretability via atom-level analysis of the dictionary, has no assumption about the class distribution (unlike Mahalanobis distance). We evaluated SUE on several NLU benchmarks (GLUE and ANLI tasks) and sentiment analysis benchmarks (Twitter, ParaDetox, and Jigsaw). In general, SUE outperforms or matches the performance of other methods. SUE performs particularly well when there is considerable uncertainty in the model, i.e., when the model lacks high precision.
In NLP, language–gender associations are commonly grounded in the author’s gender identity, inferred from their language use. However, this identity-based framing risks reinforcing stereotypes and marginalizing individuals who do not conform to normative language–gender associations. To address this, we operationalize the language–gender association as a perceived gender expression of language, focusing on how such expression is externally interpreted by humans, independent of the author’s gender identity. We present the first dataset of itskind: 5,100 human annotations of perceived gendered style—human-written texts rated on a five-point scale from very feminine to verymasculine. While perception is inherently subjective, our analysis identifies textual features associated with higher agreement among annotators: formal expressions and lower emotional intensity. Moreover, annotator demographics influence their perception: women annotators are more likely to label texts as feminine, and men and non-binary annotators as masculine. Finally, feature analysis reveals that the text’s perceived gendered style is shaped by both affective and function words, partially overlapping with known patterns of language variation across gender identities. Our findings lay the groundwork for operationalizing gendered style through human annotation, while also highlighting annotators’ subjective judgments as meaningful signals to understand perception-based concepts.
Insensitivity to semantically-preserving variations of prompts (paraphrases) is crucial for reliable behavior and real-world deployment of large language models. However, language models exhibit significant performance degradation with semantically equivalent but differently phrased prompts, and existing solutions either depend on trial-and-error prompt engineering or require computationally expensive inference-time algorithms. In this study, built on the key insight that worst-case prompts exhibit a drift in embedding space, we present Latent Adversarial Paraphrasing (LAP), a dual-loop adversarial framework that optimizes a trainable perturbation as “latent continuous paraphrase” and language model performance on these perturbations iteratively. Extensive experiments are conducted to demonstrate the effectiveness of LAP across multiple backbones on the RobustAlpaca benchmark with a 0.5%-4% absolution improvement on worst-case win-rate.
Logical reasoning is essential for large language models (LLMs) to ensure accurate and coherent inference. However, LLMs struggle with reasoning order variations and fail to generalize across logically equivalent transformations. LLMs often rely on fixed sequential patterns rather than true logical understanding. To address this issue, we introduce an order-centric data augmentation framework based on commutativity in logical reasoning. We first randomly shuffle independent premises to introduce condition order augmentation. For reasoning steps, we construct a directed acyclic graph (DAG) to model dependencies between steps, which allows us to identify valid reorderings of steps while preserving logical correctness. By leveraging order-centric augmentations, models can develop a more flexible and generalized reasoning process. Finally, we conduct extensive experiments across multiple logical reasoning benchmarks, demonstrating that our method significantly enhances LLMs’ reasoning performance and adaptability to diverse logical structures. We release our codes and augmented data in https://anonymous.4open.science/r/Order-Centric-Data-Augmentation-822C.
We introduce CopySpec, a simple yet effective technique to tackle the inefficiencies LLMs face when generating responses that closely resemble previous outputs or responses that can be verbatim extracted from context. CopySpec identifies repeated sequences in the model’s chat history or context and speculates that the same tokens will follow, enabling seamless copying without compromising output quality and without requiring additional GPU memory. To evaluate the effectiveness of our approach, we conducted experiments using seven LLMs and five datasets: MT-Bench, CNN/DM, GSM8K, HumanEval, and our newly created dataset, MT-Redundant. MT-Redundant, introduced in this paper, transforms the second turn of MT-Bench into a request for variations of the first turn’s answer, simulating real-world scenarios where users request modifications to prior responses. Our results demonstrate significant speed-ups: up to 2.35x on CNN/DM, 3.08x on the second turn of select MT-Redundant categories, and 2.66x on the third turn of GSM8K’s self-correction tasks. Importantly, we show that CopySpec integrates seamlessly with speculative decoding, yielding an average 49% additional speed-up over speculative decoding for the second turn of MT-Redundant across all eight categories. While LLMs, even with speculative decoding, suffer from slower inference as context size grows, CopySpec leverages larger contexts to accelerate inference, making it a faster complementary solution. Our code and dataset are publicly available at https://github.com/RazvanDu/CopySpec.
Natural Language Inference (NLI) is an important task in natural language processing. NLI models are aimed at automatically determining logical relationships between pairs of sentences. However, recent studies based on gold labels assigned to sentence pairs by human experts have provided some evidence that NLI models tend to make inconsistent model decisions during inference. Previous studies have used existing NLI datasets to test the transitive consistency of language models. However, they test only variations of two transitive consistency rules out of four. To further evaluate the transitive consistency of NLI models, we propose a novel evaluation approach that allows us to test all four rules automatically by generating adversarial examples via antonym replacements. Since we are testing self-consistency, human labeling of generated adversarial examples is unnecessary. Our experiments on several benchmark datasets indicate that the examples generated by the proposed antonym replacement methodology can reveal transitive inconsistencies in the state-of-the-art NLI models.
EverTracer: Hunting Stolen Large Language Models via Stealthy and Robust Probabilistic Fingerprint
PDF ↗The proliferation of large language models (LLMs) has intensified concerns over model theft and license violations, necessitating robust and stealthy ownership verification. Existing fingerprinting methods either require impractical white-box access or introduce detectable statistical anomalies. We propose EverTracer, a novel gray-box fingerprinting framework that ensures stealthy and robust model provenance tracing. EverTracer is the first to repurpose Membership Inference Attacks (MIAs) for defensive use, embedding ownership signals via memorization instead of artificial trigger-output overfitting. It consists of Fingerprint Injection, which fine-tunes the model on any natural language data without detectable artifacts, and Verification, which leverages calibrated probability variation signal to distinguish fingerprinted models. This approach remains robust against adaptive adversaries, including input level modification, and model-level modifications. Extensive experiments across architectures demonstrate EverTracer’s state-of-the-art effectiveness, stealthness, and resilience, establishing it as a practical solution for securing LLM intellectual property.
Fine-tuning large language models (LLMs) often causes overfitting to specific prompt wording, where minor phrasing variations drastically reduce performance. To address this, we propose Prompt-Agnostic Fine-Tuning (PAFT), a method that enhances robustness through dynamic prompt variation during training. PAFT first generates diverse synthetic prompts, then continuously samples from this set to construct training instances, forcing models to learn fundamental task principles rather than surface-level patterns. Across systematic evaluations using both supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RLFT), PAFT consistently demonstrates improved performance on benchmarks for question answering, mathematical reasoning, and tool use. It achieves 7% higher generalization accuracy on unseen prompts than standard methods with similar training efficiency. Notably, models trained with PAFT attain 3.2× faster inference speeds due to reduced prompt sensitivity. Ablation studies further validate effectiveness of PAFT, while theoretical analysis reveals that PAFT can effectively enhance the cross-domain generalization ability of LLM.
Optimization · Large Scale, Parallel and Distributed
Diffusion models have demonstrated excellent performance for real-world image super-resolution (Real-ISR), albeit at high computational costs. Most existing methods are trying to derive one-step diffusion models from multi-step counterparts through knowledge distillation (KD) or variational score distillation (VSD). However, these methods are limited by the capabilities of the teacher model, especially if the teacher model itself is not sufficiently strong. To tackle these issues, we propose a new One-Step \textbf{D}iffusion model with a larger-scale \textbf{D}iffusion \textbf{D}iscriminator for SR, called D$^3$SR. Our discriminator is able to distill noisy features from any time step of diffusion models in the latent space. In this way, our diffusion discriminator breaks through the potential limitations imposed by the presence of a teacher model. Additionally, we improve the perceptual loss with edge-aware DISTS (EA-DISTS) to enhance the model's ability to generate fine details. Our experiments demonstrate that, compared with previous diffusion-based methods requiring dozens or even hundreds of steps, our D$^3$SR attains comparable or even superior results in both quantitative metrics and qualitative evaluations. Moreover, compared with other methods, D$^3$SR achieves at least $3\times$ faster inference speed and reduces parameters by at least 30\%.
Computer Vision · Image and Video Generation
Diffusion Transformers (DiTs) have recently demonstrated remarkable performance in visual generation tasks, surpassing traditional U-Net-based diffusion models by significantly improving image and video generation quality and scalability. However, the large model size and iterative denoising process introduce substantial computational and memory overhead, limiting their deployment in real-world applications. Post-training quantization (PTQ) is a promising solution that compresses models and accelerates inference by converting weights and activations to low-bit representations. Despite its potential, PTQ faces significant challenges when applied to DiTs, often resulting in severe degradation of generative quality. To address these issues, we propose VETA-DiT (**V**ariance-**E**qualized and **T**emporal **A**daptation for **Di**ffusion **T**ransformers), a dedicated quantization framework for DiTs. Our method first analyzes the sources of quantization error from the perspective of inter-channel variance and introduces a Karhunen–Loève Transform enhanced alignment to equalize variance across channels, facilitating effective quantization under low bit-widths. Furthermore, to handle the temporal variation of activation distributions inherent in the iterative denoising steps of DiTs, we design an incoherence-aware adaptive method that identifies and properly calibrates timesteps with high quantization difficulty. We validate VETA-DiT on extensive image and video generation tasks, preserving acceptable visual quality under the more aggressive W4A4 configuration. Specifically, VETA-DiT reduces FID by 33.65 on the DiT-XL/2 model and by 45.76 on the PixArt-$\Sigma$ model compared to the baseline under W4A4, demonstrating its strong quantization capability and generative performance. Code is available at: https://github.com/xululi0223/VETA-DiT.
Probabilistic Methods · Bayesian Models and Methods
Estimating the uncertainty of responses from Large Language Models (LLMs) remains a critical challenge. While recent Bayesian methods have demonstrated effectiveness in quantifying uncertainty through low-rank weight updates, they typically require complex fine-tuning or post-training procedures. In this paper, we propose **T**raining-**F**ree **B**ayesianization (**TFB**), a simple yet theoretically grounded framework that efficiently transforms trained low-rank adapters into Bayesian ones without additional training. TFB systematically searches for the maximally acceptable level of variance in the weight posterior, constrained within a family of low-rank isotropic Gaussian distributions. Our theoretical analysis shows that under mild conditions, this search process is equivalent to KL-regularized variational optimization, a generalized form of variational inference. Through comprehensive experiments, we show that TFB achieves superior uncertainty estimation and generalization compared to existing methods while eliminating the need for complex Bayesianization training procedures.
Computer Vision · Image and Video Generation
Latent diffusion models (LDMs) dominate high-quality image generation, yet integrating representation learning with generative modeling remains a challenge. We introduce a novel generative image modeling framework that seamlessly bridges this gap by leveraging a diffusion model to jointly model low-level image latents (from a variational autoencoder) and high-level semantic features (from a pretrained self-supervised encoder like DINO). Our latent-semantic diffusion approach learns to generate coherent image-feature pairs from pure noise, significantly enhancing both generative quality and training efficiency, all while requiring only minimal modifications to standard Diffusion Transformer architectures. By eliminating the need for complex distillation objectives, our unified design simplifies training and unlocks a powerful new inference strategy: Representation Guidance, which leverages learned semantics to steer and refine image generation. Evaluated in both conditional and unconditional settings, our method delivers substantial improvements in image quality and training convergence speed, establishing a new direction for representation-aware generative modeling.
Optimization · Neural Network Optimization
Sparse and low-rank matrix composite approximation has emerged as a promising paradigm for compressing large language models (LLMs), offering a more flexible pruning structure than conventional methods based solely on sparse matrices. The significant variation in weight redundancy across layers, along with the differing rank and sparsity structures of weight matrices, makes identifying the globally optimal pruning structure extremely challenging. Existing methods often depend on uniform or manually designed heuristic rules to allocate weight sparsity across layers, subsequently compressing each matrix using matrix approximation techniques. Given the above theoretical difficulty in global compression of LLMs and the limited computational and data resources available compared to the training phase, we argue that a collaboration between learning and matrix approximation is essential for effective compression. In this paper, we propose a novel LLM compression framework based on generalized bilevel optimization that naturally formulates an effective collaborative mechanism. Specifically, the outer loop frames the weight allocation task as a probabilistic optimization problem, enabling the automatic learning of both layer-wise sparsities and matrix-wise retained ranks, while the inner loop solves the corresponding sparsity and rank-constrained model compression problem via matrix approximation. Our main technical contributions include two key innovations for efficiently solving this bilevel optimization problem. First, we introduce a truncated Gaussian prior-based probabilistic parameterization integrated with a policy gradient estimator, which avoids expensive backpropagation and stabilizes the optimization process. Second, we design an adapted QR-based matrix approximation algorithm that significantly accelerates inner loop computations. Extensive experiments on Phi-3 and the LLama-2/3 family demonstrate the effectiveness of our method. Notably, it maintains over 95\% zero-shot accuracy under 50\% sparsity and achieves up to 2× inference speedup.