Large Language Models (LLMs) are susceptible to indirect prompt injection attack, where the model inadvertently responds to instructions injected into the prompt context. This vulnerability stems from LLMs’ inability to distinguish between data and instructions within a prompt. We propose CachePrune that defends against this attack by identifying and pruning neurons associated with instruction-following, during KV cache encoding of the prompt context. The pruning steers the LLM toward interpreting the context purely as data rather than as instructions to follow. To identify these neurons, we introduce a neural attribution mechanism guided by a preferential attribution loss, and theoretically connect this loss to an upper bound of the Direct Preference Optimization (DPO) objective. Further, we improve on the fidelity of neural attribution by leveraging an observed triggering effect in instruction-following. Our approach does not interfere with prompt formatting or incur test-time overhead in response generation. Experiments show that CachePrune significantly reduces the attack success rate while preserving the LLM’s ability to follow user instructions.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Sarcasm detection remains a significant challenge due to its reliance on nuanced contextual understanding, world knowledge, and multi-faceted linguistic cues that vary substantially across different sarcastic expressions. Existing approaches, from fine-tuned transformers to large language models, apply a uniform reasoning strategy to all inputs, struggling to address the diverse analytical demands of sarcasm. These demands range from modeling contextual expectation violations to requiring external knowledge grounding or recognizing specific rhetorical patterns. To address this limitation, we introduce RAM-SD, a Retrieval-Augmented Multi-Agent framework for Sarcasm Detection. The framework operates through four stages: (1) contextual retrieval grounds the query in both sarcastic and non-sarcastic exemplars; (2) a meta-planner classifies the sarcasm type and selects an optimal reasoning plan from a predefined set; (3) an ensemble of specialized agents performs complementary, multi-view analysis; and (4) an integrator synthesizes these analyses into a final, interpretable judgment with a natural language explanation. Evaluated on four standard benchmarks, RAM-SD achieves a state-of-the-art Macro-F1 of 77.74%, outperforming the strong GPT-4o+CoC baseline by 7.01 points. Our framework not only sets a new performance benchmark but also provides transparent and interpretable reasoning traces, illuminating the cognitive processes behind sarcasm comprehension.
Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining
PDF ↗Large language models (LLMs) learn non-trivial abstractions during pretraining, such as detecting irregular plural noun subjects. However, because traditional evaluation methods (e.g., benchmarking) fail to reveal how models acquire these concepts and capabilities, it is not well understood when and how these specific linguistic abilities emerge. To bridge this gap and better understand model training at the concept level, we use sparse crosscoders to discover and align features across model checkpoints. Using this approach, we track the evolution of linguistic features during pretraining. We train crosscoders between open-sourced checkpoint triplets with significant performance and representation shifts, and introduce a novel metric, Relative Indirect Effects (RelIE), to trace training stages at which individual features become causally important for task performance. We show that crosscoders can detect feature emergence, maintenance, and discontinuation during pretraining. Our approach is architecture-agnostic and scalable, offering a promising path toward more interpretable and fine-grained analysis of representation learning throughout pretraining.
VALU: A Benchmark for Video Anomaly Temporal Localization and Understanding at Multiple Semantic Levels
PDF ↗Video anomaly understanding (VAU) is critical for real-world scenarios. Recent advances in Video Large Language Models (Video-LLMs) enhance the ability of VAU models to describe and interpret anomalies. However, progress in anomaly localization is still limited by two key issues. First, most existing video anomaly datasets only annotate segments that are clearly inconsistent with the context, often omitting subsequent segments that are semantically part of the same abnormal event. Second, the field lacks systematic evaluation protocols. To bridge these gaps, we introduce VALU, a new benchmark that explicitly defines anomalies across five semantic levels and provides comprehensive temporal boundaries and detailed textual descriptions for each. Based on these annotations, we design three evaluation tasks that comprehensively assess models’ capabilities across different dimensions, including temporal grounding, anomaly localization, and anomaly detail discrimination. Evaluation results reveal persistent challenges in current models’ capabilities on VAU. We further analyze and discuss these findings, and hope that both VALU and insights will advance research in VAU and the development of Video-LLMs. Our benchmark will be publicly available.
Large language models (LLMs) are shaping global values, yet they frequently exhibit a pronounced WEIRD (Western, Educated, Industrialized, Rich, Democratic) cultural bias, marginalizing diverse viewpoints and posing challenges for reconciling diverse populations with varying cultural backgrounds and value systems. In this work, we move beyond simple alignment methods to propose a new paradigm for cross-cultural fairness. We introduce a Nash Consensus Negotiation framework under the formulation of cross-cultural consensus as a Nash Equilibrium. Each LLM iteratively proposes and refines natural-language guidelines, guided by a utility function balancing self-consistency with mutual acceptance, while penalizing redundancy. The process expands the proposal space and converges to a consensus, yielding fair and interpretable consensus outcomes. We evaluate our framework against baselines using quantitative metrics, qualitative analysis, and large-scale human studies. Experiments demonstrate that our framework generates higher-quality and more balanced consensus, effectively mitigating assimilation toward WEIRD values. Furthermore, we finetune diverse LLM architectures with negotiation data via preference optimization and supervised reasoning, reducing cultural distances by up to 95.53%. Overall, our work offers a systematic path to mitigate cultural bias in LLMs by guiding them toward self-consistency, mutually-acceptable equilibria.
Despite strong results on many tasks, multimodal large language models (MLLMs) still underperform on visual mathematical problem solving, especially in reliably perceiving and interpreting diagrams. Inspired by human problem-solving, we hypothesize that the ability to extract meaningful information from diagrams is pivotal, as it directly conditions subsequent inference.Hence, we introduce FlowVerse, a comprehensive benchmark that provides a fine-grained evaluation of MLLMs’ perception and reasoning capabilities. Our preliminary results on FlowVerse reveal that existing MLLMs exhibit substantial limitations when extracting essential information and reasoned properties from diagrams and performing complex reasoning based on these visual inputs. In response, we introduce MathFlow, a modular problem-solving pipeline that decouples perception and inference into distinct stages, thereby optimizing each independently. Given the perceptual limitations observed in current MLLMs, we trained MathFlow-P-7B as a dedicated perception model.Experimental results indicate that MathFlow-P-7B yields substantial performance gains when integrated with various closed-source and open-source inference models. This demonstrates the effectiveness of the MathFlow pipeline and its compatibility with diverse inference frameworks. Project page: https://github.com/MathFlow-zju/MathFlow.
Large language models (LLM) have emerged as a promising avenue for time series forecasting, offering the potential to integrate multimodal data. However, existing LLM-based approaches face notable limitations—such as marginalized role in model architectures, reliance on coarse statistical text prompts, and lack of interpretability. In this work, we introduce Augur, a fully LLM driven time series forecasting framework that exploits LLM causal reasoning to discover and use directed causal associations among covariates. Augur uses a two stage teacher student architecture where a powerful teacher LLM infers a directed causal graph from time series using heuristic search together with pairwise causality testing. A lightweight student agent then refines the graph and fine tune on high confidence causal associations that are encoded as rich textual prompts to perform forecasting. This design improves predictive accuracy while yielding transparent, traceable reasoning about variable interactions. Extensive experiments on real-world datasets with 25 baselines demonstrate that Augur achieves competitive performance and robust zero-shot generalization.
Deep Learning · Generative Models and Autoencoders
Continuous Normalizing Flows (CNFs) enable elegant generative modeling but remain bottlenecked by their iterative nature requiring costly sampling and lacking interpretability of the intermediate states. Recent approaches accelerate sampling by straightening trajectories or distilling endpoints, yet they treat the original generative process as a black box, discarding the teacher’s intermediate dynamics. We propose a fundamentally different perspective: globally linearizing flow dynamics via Koopman theory to achieve trajectory-preserving linearization. By lifting Conditional Flow Matching (CFM) into a higher-dimensional Koopman space, we represent its evolution with a single linear operator. Crucially, unlike boundary-only distillation, our method enforces infinitesimal consistency with the teacher's vector field along *the full generative path*. We derive a practical, simulation-free training objective that ensures this global alignment and yields two key benefits. First, sampling becomes one-step and parallelizable. Second, because the linearization is faithful to the dynamics, the Koopman operator provides unique insights on the generation. We demonstrate that this structure enables novel applications unavailable in prior approaches, including discovery of semantically coherent editing directions, inversion with a teacher-aligned linear operator and class-conditional spectral signatures. Empirically, our approach achieves competitive sample quality, while enabling spectral analysis and control of the *entire trajectories* of generative flows.
Social Aspects · Accountability, Transparency, and Interpretability
Modern AI systems are typically developed through multiple stages-pretraining, fine-tuning rounds, and subsequent adaptation or alignment, where each stage builds on the previous ones and updates the model in distinct ways. This raises a critical question of accountability: when a deployed model succeeds or fails, which stage is responsible, and to what extent? We pose the accountability attribution problem for tracing model behavior back to specific stages of the model development process. To address this challenge, we propose a general framework that answers counterfactual questions about stage effects: how would the model's behavior have changed if the updates from a particular stage had not occurred? Within this framework, we introduce estimators that efficiently quantify stage effects without retraining the model, accounting for both the data and key aspects of model optimization dynamics, including learning rate schedules, momentum, and weight decay. We demonstrate that our approach successfully quantifies the accountability of each stage to the model's behavior. Based on the attribution results, our method can identify and remove spurious correlations learned during image classification and text toxicity detection tasks that were developed across multiple stages. Our approach provides a practical tool for model analysis and represents a significant step toward more accountable AI development.
Applications · Everything Else
While Transformer-based architectures have revolutionized neural decoding, existing models often treat codes as generic sequences, ignoring their inherent algebraic properties. In this paper, we take a step toward bridging these two domains by proposing a novel decoding approach that integrates the algebraic structure of cyclic codes into Transformer-based decoders. Leveraging the inherent cyclic properties, we introduce interpretable error correction patterns and inter-node relationship hypotheses that link the structural characteristics of the codes to the model parameters. Building on these insights, we design a plug-and-play, flexibly deployable decoding method tailored for cyclic codes. Experimental results show that our method achieves an average reduction in bit error rate (BER) by an order of magnitude, while also reducing the total number of parameters by approximately 97%. Additional comparative experiments validate our proposed conjectures and highlight a promising pathway for bridging classical coding theory and modern Transformer-based decoding architectures.
General Machine Learning · Representation Learning
Determining the complexity, or Intrinsic Dimension (ID), of data is fundamental to efficient and interpretable representation learning. This is particularly challenging in multi-modal settings when trying to learn disentangled representations for shared and private information. Existing techniques leave a critical gap: they are often static, uni-modal, or in the case of contrastive methods, adapt only to the shared ID implicitly. We introduce Fidelity-Guided Rank Optimization (FiGuRO), a framework for learning the complete ID structure of uni- and multi-modal data. FiGuRO learns the dimensions of low-rank projections using truncated singular value decomposition and an algorithm that determines when to reduce or increase dimensionalities and in which latent spaces. Disentanglement of shared and private information arises as an emergent property of this optimization, eliminating the need for complex auxiliary loss functions. We demonstrate that FiGuRO outperforms existing ID estimation techniques and is more robust to hyperparameter changes. Across simulations and real-world data, FiGuRO captures distinct ID scales and varying subspace ratios, and decomposes shared and private information successfully. Furthermore, we show that FiGuRO can be applied to modern uni-modal pretrained models, enabling efficient, post-hoc disentanglement of multi-modal representations.
Multi-Hop Fact Verification (MHFV) necessitates complex reasoning across disparate evidence, posing significant challenges for Large Language Models (LLMs) which often suffer from hallucinations and fractured logical chains. Existing methods, while improving transparency via Chain-of-Thought (CoT), lack explicit modeling of the causal dependencies between evidence and claims. In this work, we introduce a novel framework that grounds reasoning in a Structural Causal Model (SCM), treating verification as a constructive causal inference process. We empirically identify an "inverted U-shaped" correlation between reasoning chain length and accuracy, revealing that excessive structural complexity degrades performance. To address this, we propose a Rule-based Reinforcement Learning strategy using Group Relative Policy Optimization (GRPO). This approach dynamically optimizes the trade-off between structural depth and conciseness. Extensive experiments on HoVer and EX-FEVER demonstrate that our SCM-GRPO framework significantly outperforms state-of-the-art baselines, offering a reliable and interpretable solution for complex fact verification.
Deep Learning · Large Language Models
Large language models (LLMs) are increasingly adapted to downstream tasks in resource-constrained scenarios, making quantization-aware fine-tuning (QAF) a common practice for practical deployment. However, we find that quantized LLMs are substantially more vulnerable to safety alignment degradation during fine-tuning than full-precision models by interpretability analyses. In this paper, we first theoretically reveal that this vulnerability is driven by quantization errors, manifesting as an initial safety shift followed by a distorted optimization path. Based on this insight, we propose Explicit-Safety Quantization-Aware Fine-tuning (ExSQF), which effectively restores model safety while preserving downstream performance. It initializes adapters by combining quantization error with a safety matrix projection to mitigate early safety shifts, followed by post-training refinement that corrects deviations in the optimization path. Extensive experimental results show that ExSQF achieves state-of-the-art safety alignment recovery, even surpassing existing full-precision safety-aware fine-tuning baseline, while effectively preserving model performance.
Applications · Chemistry, Physics, and Earth Sciences
Climate change research increasingly requires AI systems that reason across text, dynamic visual content, and scientific figures, yet existing climate QA benchmarks are small, mostly textual, and cover a narrow range of models. We introduce MMClima, a large-scale multimodal climate question answering framework with over 104k expert-validated question–answer pairs spanning articles, video transcriptions, and figures across five core climate science domains. MMClima is constructed via automated claim extraction and QA synthesis with human-in-the-loop validation to ensure both scale and reliability. Using MMClima, we benchmark state-of-the-art multimodal language models on tasks requiring factual recall, visual interpretation, and cross-modal synthesis. We additionally fine-tune on the textual split to produce mmclima-70b-txt, a domain-adapted baseline that outperforms strong open- and closed-source models on textual QA. We release the dataset, evaluation pipeline, fine-tuned model weights, and data creation framework to support standardized multimodal evaluation for climate science.
Applications · Neuroscience, Cognitive Science
Foundation models for intracranial neural recordings aim to learn generalizable representations from large-scale unlabeled data. However, existing approaches rely on suboptimal tokenization schemes -- treating individual electrode channels as independent tokens or aggregating them into a single brain-wide representation -- which fail to capture the brain’s inherent functional modularity. We introduce NeuroCLUS, a foundation model that learns to represent neural activity through data-driven functional clusters. NeuroCLUS is built on a novel two-stage pre-training framework. First, a spatial-temporal model learns a functional context graph between channels via a functional context prediction task. Second, this graph guides a soft clustering of channels into a set of learnable prototype tokens, enabling the transformer backbone to process coherent functional units rather than raw channels. Evaluated across a diverse range of decoding paradigms -- including speech perception, speech production, and seizure detection -- NeuroCLUS consistently achieves state-of-the-art performance. The discovered functional clusters align with established neurophysiology and offer enhanced interpretability. Our work demonstrates that explicitly modeling functional neural groupings significantly improves the efficiency, generalization, and interpretability of foundation models for intracranial decoding.
Deep Learning · Sequential Models, Time series
Representing the past in a compressed, efficient, and informative manner is a central problem for systems trained on sequential data. The *HiPPO* framework, originally proposed by Gu & Dao et al., provides a principled approach to sequential compression by projecting signals onto orthogonal polynomial (OP) bases via structured linear ordinary differential equations. Subsequent works have embedded these dynamics in state space models (SSMs), where HiPPO structure serves as an initialization. Nonlinear successors of these SSM methods such as Mamba are state of the art for many tasks with long-range dependencies, but the mechanisms by which they represent and prioritize history remain largely implicit. In this work, we revisit the HiPPO framework with the goal of making these mechanisms explicit. We show how polynomial representations of history can be extended to support capabilities of modern SSMs such as adaptive allocation of memory, and input-dependent state updates, and associative memory, while retaining direct interpretability in the OP basis. We introduce a unified framework comprising five such extensions, which we collectively refer to as a "HiPPO zoo." Each extension exposes a specific modeling capability as an explicit modification of the underlying measure or dynamics governing the polynomial coefficients, rather than as an opaque learned transformation. The resulting models adapt their memory online and train in streaming settings with efficient updates. We illustrate the behaviors and advantages of these extensions through a range of synthetic sequence modeling tasks, highlighting how explicit polynomial memories can recover and clarify mechanisms implicit in SSMs.
Zero-shot out-of-vocabulary detection (ZS-OOVD) aims to accurately recognize objects of in-vocabulary (IV) categories provided at zero-shot inference, while simultaneously rejecting undefined ones (out-of-vocabulary, OOV) that lack corresponding category prompts. However, previous methods are prone to overfitting the IV classes, leading to the OOV or undefined classes being misclassified as IV ones with a high confidence score. To address this issue, this paper proposes a zero-shot OOV detector (OOVDet), a novel framework that effectively detects predefined classes while reliably rejecting undefined ones in zero-shot scenes. Specifically, due to the model’s lack of prior knowledge about the distribution of OOV data, we synthesize region-level OOV prompts by sampling from the low-likelihood regions of the class-conditional Gaussian distributions in the hidden space, motivated by the assumption that unknown semantics are more likely to emerge in low-density areas of the latent space. For OOV images, we further propose a Dirichlet-based gradient attribution mechanism to mine pseudo-OOV image samples, where the attribution gradients are interpreted as Dirichlet evidence to estimate prediction uncertainty, and samples with high uncertainty are selected as pseudo-OOV images. Building on these synthesized OOV prompts and pseudo-OOV images, we construct the OOV decision boundary through a low-density prior constraint, which regularizes the optimization of OOV classes using Gaussian kernel density estimation in accordance with the above assumption. Experimental results show that our method significantly improves the OOV detection performance in zero-shot scenes. Code will be available.
We study the population loss landscape of two-layer ReLU networks of the form $\sum_{k=1}^K \mathrm{ReLU}(w_k^\top x)$ in a realisable teacher–student setting with Gaussian covariates. We show that local minima admit an exact low-dimensional representation in terms of \emph{summary statistics}, yielding a sharp and interpretable characterisation of the landscape. We further establish a direct link with one-pass SGD: local minima correspond to attractive fixed points of the dynamics in summary statistics space. This perspective reveals a hierarchical structure of minima: they are typically isolated in the well-specified regime, but become connected by flat directions as network width increases. In this overparameterised regime, global minima become increasingly accessible, attracting the dynamics and reducing convergence to spurious solutions. Overall, our results reveal intrinsic limitations of common simplifying assumptions, which may miss essential features of the loss landscape even in minimal neural network models.
Deep Learning · Large Language Models
Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model's performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model's geometry, as captured by existing metrics, primarily reflects training choices rather than performance.
Reinforcement Learning · Everything Else
Controlling autonomous systems under real-world conditions often requires policies that can be evaluated with low latency and low energy requirements. Unfortunately, these conditions are at odds with the use of high-precision deep networks as controllers. In this work, we introduce Differentiable Weightless Controllers (DWCs), a symbolic-differentiable architecture that allows learning flexible non-linear yet highly efficient control policies. DWCs can be trained end-to-end by gradient-based techniques, yet compile directly into FPGA-compatible circuits with few- or even single-clock-cycle latency and nanojoule-level energy cost per action for the core computation. Across five MuJoCo benchmarks, including high-dimensional Humanoid, DWCs achieve returns competitive with standard deep policies (full precision or quantized neural networks). Furthermore, DWCs exhibit structurally sparse and interpretable connectivity patterns, enabling a direct inspection of which input values influence control decisions.