论文检索

Deep Learning · Attention Mechanisms

Difan Deng, Andreas B. Winje, Lukas Fehring, Marius Lindauer

The quadratic computational complexity of softmax transformers has become a bottleneck in long-context scenarios. In contrast, linear attention model families provide a promising direction towards a more efficient sequential model. These linear attention models compress past $KV$ values into a single hidden state, thereby efficiently reducing complexity during both training and inference. However, their expressivity remains limited by the size of their hidden state. Previous work proposed interleaving softmax and linear attention layers to reduce computational complexity while preserving expressivity. Nevertheless, the efficiency of these models remains bottlenecked by their softmax attention layers. In this paper, we propose Neural Attention Search Linear (NAtS-L), a framework that applies both linear attention and softmax attention operations within the same layer on different tokens. NAtS-L automatically determines whether a token can be handled by a linear attention model, i.e., tokens that have only short-term impact and can be encoded into fixed-size hidden states, or require softmax attention, i.e., tokens that contain information related to long-term retrieval and need to be preserved for future queries. By searching for optimal Gated DeltaNet and softmax attention combinations across tokens, we show that NAtS-L provides a strong yet efficient token-level hybrid architecture.

Social Aspects · Accountability, Transparency, and Interpretability

Khalid Oublal, Quentin Bouniot, Qi Gan, Stephan Clemencon, Zeynep Akata

As black box models and pretrained models gain traction in time series applications, understanding and explaining their predictions becomes increasingly vital, especially in high-stakes domains where interpretability and trust are essential. However, most of the existing methods involve only in-distribution explanation, and do not generalize outside the training support, which requires the learning capability of generalization. In this work, we aim to provide a framework to explain black-box models for time series data through the dual lenses of Sparse Autoencoders (SAEs) and causality. We show that many current explanation methods are sensitive to distributional shifts, limiting their effectiveness in real-world scenarios. Building on the concept of Sparse Autoencoder, we introduce TimeSAE, a framework for black-box model explanation. We conduct extensive evaluations of TimeSAE on both synthetic and real-world time series datasets, comparing it to leading baselines. The results, supported by both quantitative metrics and qualitative insights, show that TimeSAE provides more faithful and robust explanations. Our code is available in an easy-to-use library TimeSAE-Lib: https://anonymous.4open.science/w/TimeSAE-571D/

General Machine Learning · Unsupervised and Semi-supervised Learning

Simon Roschmann, Paul KRZAKALA, Sonia Mazelet, Quentin Bouniot, Zeynep Akata

The Platonic Representation Hypothesis posits that neural networks trained on different modalities converge toward a shared statistical model of the world. Recent work exploits this convergence by aligning frozen pretrained vision and language models with lightweight alignment layers, but typically relies on contrastive losses and millions of paired samples. In this work, we ask whether meaningful alignment can be achieved with substantially less supervision. We introduce a semi-supervised setting in which pretrained unimodal encoders are aligned using a small number of image–text pairs together with large amounts of unpaired data. To address this challenge, we propose SOTAlign, a two-stage framework that first recovers a coarse shared geometry from limited paired data using a linear teacher, then refines the alignment on unpaired samples via an optimal-transport-based divergence that transfers relational structure without overconstraining the target space. Unlike existing semi-supervised methods, SOTAlign effectively leverages unpaired images and text, learning robust joint embeddings across datasets and encoder pairs, and significantly outperforming supervised and semi-supervised baselines.

Deep Learning · Generative Models and Autoencoders

Mateusz Pach, Jessica Bader, Quentin Bouniot, Serge Belongie, Zeynep Akata

Text-to-image generation models have advanced rapidly, yet achieving fine-grained control over generated images remains difficult, largely due to limited understanding of how semantic information is encoded. We develop an interpretation of the color representation in the Variational Autoencoder latent space of FLUX.1 [Dev], revealing a structure reflecting Hue, Saturation, and Lightness. We verify our Latent Color Subspace (LCS) interpretation by demonstrating that it can both predict and explicitly control color, introducing a fully training-free method in FLUX based solely on closed-form latent-space manipulation.

Deep Learning · Foundation Models

Yiran Huang, Karsten Roth, Quentin Bouniot, Wenjia Xu, Zeynep Akata

Transformer-based multimodal large language models often exhibit in-context learning (ICL) capabilities. Motivated by this phenomenon, we ask: how do transformers learn to associate information across modalities from in-context examples? We investigate this through controlled experiments on small transformers trained on synthetic classification tasks, enabling precise manipulation of data statistics and model architecture. We begin by revisiting core principles of unimodal ICL in modern transformers. While several prior findings replicate, we find that Rotary Position Embeddings (RoPE) can delay the onset of ICL circuits. Extending to the multimodal setting reveals a fundamental learning asymmetry: when pretrained on high-diversity data from a primary modality, surprisingly low data complexity in the secondary modality suffices for multimodal ICL to emerge. Mechanistic analysis shows that both settings rely on an induction-style mechanism that copies labels from matching in-context exemplars; multimodal training refines and extends these circuits across modalities. Our findings provide a mechanistic foundation for understanding multimodal ICL in modern transformers and introduce a controlled testbed for future investigation.

Theory · Everything Else

Alireza Jafari, Heman Shakeri, Hadi Daneshmand

Single-cell perturbation analysis aims to predict how cellular states change after interventions such as drug treatments or genetic edits. A central difficulty is that pre- and post-perturbation measurements are typically observed as *unpaired* populations, so accurate prediction requires inferring a latent coupling and learning a transition map. In this position paper, we argue that the *measurement time gap* is the key experimental knob controlling both the computational tractability of coupling and the effective model complexity. We identify a critical time gap $\Delta$ that induces a phase transition, under biologically inspired conditions; for "measurement-time $< \Delta$", matching is polynomial-time tractable and the task reduces to supervised learning, whereas for "measurement-time $>\Delta$", recovering the matching is NP-hard in the worst case. The required conditions are restricted isometry of the initial states and temporal smoothness of the transition dynamics. We complement the theory with empirical evidence on synthetic and biological datasets showing a sharp regime change as the time gap increases. Furthermore, we demonstrate that a linear model can match or exceed the performance of higher-capacity neural approaches when our conditions hold.

Deep Learning · Large Language Models

Canbin Huang, Tianyuan Shi, Xiaojun Quan, Jingang Wang, Jianfei Zhang, Qifan Wang

Model merging has emerged as a cost-effective approach for consolidating the capabilities of multiple LLMs without retraining. However, existing merging techniques, largely based on linear parameter arithmetic or optimization, struggle when applied to Mixture-of-Experts (MoE) architectures. We identify a critical failure mode in MoE merging, termed *routing breakdown*, in which the merged router fails to dispatch tokens to suitable experts. Routing breakdown stems from the sensitivity of the non-linear softmax and discrete Top-$k$ routing mechanisms to parameter perturbations from merging, a sensitivity further amplified by load-balancing constraints imposed during MoE pretraining. Because fine-tuned experts exhibit distinct specializations, even modest misrouting can cause severe performance degradation. To address this issue, we propose Hessian-Aware Router Calibration (HARC), a training-free framework that leverages second-order curvature information to realign the merged router. This approach admits a closed-form solution that can be efficiently solved using a matrix-free conjugate gradient method. Experiments on mathematical reasoning and code generation tasks show that HARC effectively mitigates routing breakdown across diverse MoE merging baselines and leads to substantial performance improvements.

Fan Feng, Yujia Zheng, Minghao Fu, Yongqiang Chen, Guangyi Chen, Kevin Murphy, Biwei Huang, Kun Zhang

Learning and planning in imagination using world models provides an effective paradigm for training agents for decision-making. However, existing approaches often rely on high-dimensional latent spaces or generic visual embeddings that retain many factors irrelevant to control, limiting efficiency and generalization across tasks. To this end, we study how agents can learn world models with representations that are task-specific, minimal, and sufficient for decision making. We achieve this via a closed-loop synergy between the agent and the world model, in which structured world-model learning distills task-sufficient representations from informative interaction data. On the agent side, agents perform active probing of the environment to collect informative trajectories that expose task-relevant latent factors, guided by an adaptive curriculum. On the world-model side, we learn structured representations over observations to distill compact, task-sufficient latent states from the collected interaction data. This synergy enables the recovery of task-sufficient latent representations that capture all control-relevant factors empirically. Leveraging these representations, the resulting policies achieve improved sample efficiency and systematic generalization, including generalization across skills, object–skill compositions, and previously unseen tasks on standard continuous control and robotic manipulation benchmarks.

Social Aspects · Safety

Jaeho Kim, Seokhyun Lee, Jieun Lee, Changhee Lee

AI safety research has mainly focused on two areas: technical alignment (ensuring AI systems produce human-aligned outputs) and the regulation of generative AI's societal impacts (including unemployment risk and labor market disruption). However, an equally important dimension remains underexplored: the risk inherent in dependence on AI systems themselves. In this position paper, we argue that AI safety research should address $\textbf{\textit{AI Lock-In}}$, the phenomenon whereby excessive reliance on AI systems leads to human deskilling, diminishes human capacity for independent functioning, and creates systemic vulnerabilities when AI systems become unavailable or compromised. We highlight that AI Lock-In is a systemic threat that is already emerging at individual, societal, and national levels, one that could be dramatically amplified by AI service disruptions or geopolitical conflicts. Drawing on detailed scenarios, we investigate how AI Lock-In emerges and escalates across multiple levels, ranging from individual skill atrophy to national-scale infrastructure failures. To address this, we provide guidance on how such risks can be mitigated and prepared for at each level. We contend that proactively addressing AI Lock-In before such dependencies become entrenched and irreversible is essential for preserving individual autonomy and national security.

Social Aspects · Safety

Jason Lucas, Pureheart Ogheneogaga Irikefe, Adaku Uchendu, Umniya Najaer, Cornelius Adejoro, Patrice Sterling, Dongwon Lee

Large language models are deployed globally as universal systems, yet their safety mechanisms remain English-optimized. This creates a Dual Curse for speakers of low-resource languages: a Harmfulness Curse where harmful content generation rises from 1\% in English to 35\% in languages like Hausa, Igbo, and Javanese, and a Relevance Curse where instruction-following drops by 20 percentage points, making these systems simultaneously more dangerous and less useful. Drawing on a PRISMA-guided systematic review of 207 studies, we demonstrate that this disparity stems from a pre-training bottleneck: reward models achieve only 49--50\% accuracy in low-resource languages (equivalent to random chance), rendering post-hoc alignment structurally ineffective. These technical failures become governance hazards when at least 22 countries mandate automated content moderation, creating an infrastructure that is exploitable for censorship. Therefore, we propose a socio-technical framework addressing this inequity: (1) safety context distillation during pre-training (achieving 78--89\% harm reduction); (2) participatory harm specification by affected communities; and (3) evaluation metrics jointly tracking attack resistance and false refusal rates across languages.

Social Aspects · Fairness

Pengqi Liu, Zijun Yu, Mouloud Belbahri, Arthur Charpentier, Masoud Asgharian, Jesse Cresswell

Conformal prediction (CP) offers distribution-free uncertainty quantification for machine learning models, yet its interplay with fairness in downstream decision-making remains underexplored. Moving beyond CP as a standalone operation (procedural fairness), we analyze the holistic decision-making pipeline to evaluate substantive fairness—the equity of downstream outcomes. Theoretically, we derive an upper bound that decomposes prediction-set size disparity into interpretable components, clarifying how label-clustered CP helps control method-driven contributions to unfairness. To facilitate scalable empirical analysis, we introduce an LLM-in-the-loop evaluator that approximates human assessment of substantive fairness across diverse modalities. Our experiments reveal that label-clustered CP variants consistently deliver superior substantive fairness. Finally, we empirically show that equalized set sizes, rather than coverage, strongly correlate with improved substantive fairness, enabling practitioners to design more fair CP systems.

Theory · Everything Else

Behrooz Tahmasebi, Melanie Weber

Dynamical systems model trajectory data generated by an underlying fixed dynamics, with applications ranging from biological systems to flows in physics. The identification problem concerns recovering the parameters of a system from observed trajectories. In many scientific settings, however, dynamical systems are not generic and instead exhibit symmetries imposed by physical laws, formalized as equivariance with respect to a group action. In this work, we study \emph{adaptive symmetry discovery} for dynamical system identification and address how a system can be identified from a single trajectory when it is equivariant with respect to an unknown symmetry group. To this end, we first show that for known symmetries, the system can be identified from a significantly shorter single trajectory than in the generic setting, and we precisely characterize this improvement. We then consider the automatic symmetry discovery setting, proposing a method to learn the symmetry group directly from a single trajectory and incorporate it into the identification procedure, achieving the same optimal trajectory length as in the known-symmetry case. Our analysis relies on tools from group representation theory and the expander properties of Cayley graphs, and may be of independent interest for the study of symmetries in dynamical systems.

Reinforcement Learning · Deep RL

Shiping Gao, Hongzhan Chen, Xiaojun Quan, Qifan Wang, Lifu Huang

Process reward models (PRMs) provide fine-grained reward signals along the reasoning process, improving credit assignment beyond outcome-only rewards. Training reliable PRMs often relies on step annotations or heavy verification pipelines, making them expensive to scale and refresh during online RL. Implicit PRMs mitigate this cost by learning decomposable token/step rewards from trajectory-level outcome labels, but they suffer a train-inference mismatch: training only constrains a sequence-level aggregate, while deployment queries token-level scores as local step quality. As a result, token credits are weakly identified and can become predictive of final success without faithfully reflecting which step is correct. This unreliability can even undermine a key promise of implicit PRMs—scoring many candidate tokens—because noisy per-token advantages may systematically reinforce incorrect continuations. We address this with a novel Implicit Prefix-Value Reward Model (IPVRM), which directly learns a prefix-conditioned value function estimating the probability of eventual correctness, and derives step signals via temporal-difference (TD) differences. IPVRM substantially improves step-verification F1 on ProcessBench. Building on these calibrated signals, we further propose Distribution-Level RL (DistRL), which computes TD advantages not only for sampled tokens but also for high-probability candidate tokens across the vocabulary, enabling dense counterfactual updates without additional rollouts. DistRL brings limited benefits with mis-calibrated implicit rewards, but consistently improves downstream reasoning once powered by IPVRM’s reliable prefix values.

General Machine Learning · Transfer, Multitask and Meta-learning

Vincent Herrmann, Jürgen Schmidhuber

This position paper argues that truly open-ended intelligence is bottlenecked by the challenge of *interestingness*: the ability to prospectively identify which tasks or data hold the potential for future progress. We formalize interestingness as an inductive heuristic for future compression progress and investigate its predictability using tools from Kolmogorov Complexity and Algorithmic Statistics. By analyzing complexity-runtime profiles under various priors over computable objects, we demonstrate that the *inductive property of interestingness*—the capacity for past compression progress to signal future discovery—is theoretically viable. However, we show that this property is highly sensitive to the underlying distribution of objects. We conclude by calling for a move beyond human-in-the-loop filtering or data creation, and a shift toward *introspective* models that can explicitly assess their own potential for insight. Furthermore, we advocate the engineering of scale-free synthetic environments, providing a principled roadmap for the development of truly autonomous open-ended systems.

Deep Learning · Large Language Models

Wenhao SUN, Rong-Cheng Tu, Yifu Ding, Zhao Jin, Jingyi Liao, Yongcheng Jing, Dacheng Tao

While Diffusion Language Models (DLMs) offer a flexible, arbitrary-order alternative to the autoregressive paradigm, their non-causal nature precludes standard KV caching, forcing costly hidden state recomputation at every decoding step. Existing caching approaches reduce this cost by selective hidden state updates; however, they are still limited by (i) computationally costly token-wise update identification heuristics and (ii) rigid, uniform budget allocation that fails to account for heterogeneous hidden- tate dynamics. To address these challenges, we present SPA-Cache that jointly optimizes update identification and budget allocation. First, we derive a low-dimensional singular proxy that enables the identification of update-critical tokens in a low-dimensional subspace, substantially reducing the overhead of update identification. Second, motivated by the layer-wise heterogeneity in hidden state dynamics, we introduce an adaptive strategy that allocates fewer updates to stable layers without degrading generation quality. Together, these contributions significantly improve the efficiency of DLMs, yielding up to an $8\times$ throughput improvement over vanilla models decoding and a $2$-$4\times$ speedup over existing caching baselines.

Deep Learning · Large Language Models

Shrey Shah, Justin Wagle

Multi-adapter serving systems route entire sequences to a single adapter, forcing a choice when requests span multiple domains. This assumption fails in two important settings: (1) multimodal generation, where text and image tokens require different adapters within the same sequence, and (2) mixed-capability requests like ``write code to solve this equation,'' which need expertise from multiple specialized adapters. We introduce per-token routing, which routes individual tokens to adapters based on either vocabulary structure (for multimodal models) or learned gating (for semantic specialization). Per-token routing is provably optimal, achieving work $N$ for $N$ tokens versus $K \cdot N$ for per-sequence routing with $K$ adapter types. Our key contribution is MoLoRA (Mixture of LoRA), which enables composable specialization: load multiple domain-specific adapters and let a learned router select the appropriate adapter per-token. We demonstrate that specialization dramatically beats scale: MoLoRA enables Qwen3-1.7B to exceed Qwen3-8B across four reasoning benchmarks while being 4.7$\times$ smaller. This enables modular expertise at inference time: train focused LoRAs independently, combine them without retraining, and add new capabilities by simply loading new adapters.

Deep Learning · Large Language Models

Ajay Jaiswal, Lauren Hannah, Han-Byul Kim, Duc Hoang, Arnav Kundu, Mehrdad Farajtabar, Minsik Cho

Understanding how transformer components operate in LLMs is important, as it is at the core of recent technological advances in artificial intelligence. In this work, we revisit the challenges associated with interpretability of feed-forward modules (FFNs) and propose MemoryLLM, which aims to decouple FFNs from self-attention and enables us to study the decoupled FFNs as context-free token-wise neural retrieval memory. In detail, we investigate how input tokens access memory locations within FFN parameters and the importance of FFN memory across different downstream tasks. MemoryLLM achieves context-free FFNs by training them in isolation from self-attention directly using the token embeddings. This approach allows FFNs to be pre-computed as token-wise lookups (ToLs), enabling on-demand transfer between VRAM and storage, additionally enhancing inference efficiency. We also introduce Flex-MemoryLLM, positioning it between a conventional transformer design and MemoryLLM. This architecture bridges the performance gap caused by training FFNs with context-free token-wise embeddings.

General Machine Learning · Evaluation

Yunze Xiao, Gordon Dai, Shahan Ali Memon, Jen-Tse Huang, Maarten Sap, Mona Diab

In this position paper, we argue that for AI systems, ``welfare'' is a choice in mechanism and evaluation, rather than an empirically discoverable property, because welfare assessment lacks an external validation channel: there is no independent, intervention-based test that can falsify a welfare metric or adjudicate among competing accounts of what welfare requires. We formalize this diagnosis using evaluation theory, emphasizing that in AI the subject, indicators, and metrics are co-engineered, so proposed welfare evidence can be manufactured or suppressed by ordinary development decisions. We then analyze two institutional failure modes if welfare scorecards are nonetheless used in release and access decisions: they expand procedural gates around routine ML work and they enable organizations to reframe discretionary choices about liability, publicity, and risk posture as moral necessity. We conclude with guidance for research and governance: prohibit welfare scorecards as release gates, disallow appeals to model welfare as a reason to resist auditing and oversight, and require that any restrictions on AI development be justified by externally verifiable harms rather than untestable welfare claims.

Mengxuan Wang, Yuxin Chen, Gang Xu, Tao He, hongjie jiang, Ming Li

Vision language models (VLMs) extend the reasoning capabilities of large language models (LLMs) to cross-modal settings, yet remain highly vulnerable to multimodal jailbreak attacks. Existing defenses predominantly rely on safety fine-tuning or \textit{aggressive} token manipulations, incurring substantial training costs or significantly degrading utility. Recent research shows that LLMs inherently recognize unsafe content in text, and the incorporation of visual inputs in VLMs frequently dilutes risk-related signals. Motivated by this, we propose Risk Awareness Injection (RAI), a \textit{lightweight} and training-free framework for safety calibration that restores LLM-like risk recognition by amplifying unsafe signals in VLMs. Specifically, RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals within the cross-modal feature space. This modulation restores the model’s LLM-like ability to detect unsafe content from visual inputs, while preserving the semantic integrity of original tokens for cross-modal reasoning. Extensive experiments across multiple jailbreak and utility benchmarks demonstrate that RAI substantially reduces attack success rate without compromising task performance.

Deep Learning · Large Language Models

Jinyang Zhang, Hongxin Ding, Yue Fang, Weibin Liao, Muyang Ye, Junfeng Zhao, Yasha Wang

Recent work has sought to understand Large Language Models (LLMs) reasoning, yet a principled, model-intrinsic signal that captures its *layer-wise reasoning dynamics* remains underexplored. We bridge this gap by demonstrating that **the $\ell_2$ norm of hidden states serves as an endogenous signal of the model's reasoning intensity**. Using Sparse Autoencoders (SAEs) as a diagnostic probe, we observe that LLMs' internal reasoning is marked by a sharp increase in reasoning feature activations concentrated in late layers. Motivated by this pattern, we establish a formal link between reasoning intensity and the model's latent geometry and theoretically prove that the $\ell_2$ norm of hidden states bounds the activation strength of SAE reasoning features. Empirical correlation analysis and causal interventions further prove $\ell_2$ norm as a faithful indicator, where heightened norms consistently correspond to critical reasoning steps. We then introduce three test-time scaling techniques guided by $\ell_2$ norms: Adaptive Layer-wise Reasoning Recursion, (ii) Endogenous Reasoning State Steering, and (iii) $\ell_2$-guided Response Selection, which requires no additional training or data and is compatible with advanced inference engines. Experiments across model architectures and benchmarks show that $\ell_2$-norm-based techniques significantly improve reasoning performance, offering a principled yet simple lens to perceive and control LLM latent reasoning dynamics. Our codes are anonymously available at https://anonymous.4open.science/r/The-Tell-Tale-Norm-4E40