What would it mean to have a *scientific* understanding of AI? Language models are not static objects—they are snapshots of time-evolving processes shaped by data, objectives, and optimization dynamics. Yet the field predominantly treats models as fixed artifacts, analyzing behaviors after training rather than asking *why* they emerge. **This position paper argues that AI research should move beyond *post hoc* fixes and study the learning dynamics of models.** We envision a hierarchy of scientific maturity: first *predict* outcomes from early training signals, then *intervene* when trajectories go wrong, ultimately *design* training procedures that guarantee desired properties. Scaling laws have reached the first level for loss; the challenge is extending all three levels to general capabilities, biases, and safety. We articulate requirements for such theories, survey progress across mechanistic interpretability, fairness, memorization, and learning dynamics, and identify concrete open problems. The path forward requires treating models as processes to be understood, not just artifacts to be patched.
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
Deep Learning · Large Language Models
The Plan-then-Code paradigm effectively enhances Large Language Models (LLMs) in complex code generation by decomposing reasoning into explicit, interpretable steps. However, introducing the plan and verification report substantially enlarges the context, which in turn misdirects the model’s attention toward irrelevant tokens and the most recently generated code. This effect leads the model to overlook critical constraints and to generate incorrect code, especially for small-scale LLMs (less than 8B). To address this issue, we propose \textbf{P}erturbation-Verified \textbf{A}ttention \textbf{D}istillation and Dynamic \textbf{A}lignment (PADA). PADA identifies the key tokens most critical to the student model and constructs the optimal attention target matrix, dynamically aligning the student’s focus with key tokens for each plan step. We evaluate PADA with two teacher models and three student models across seven benchmarks, and the results show that PADA improves Pass@1 by up to 16.7\% and outperforms SOTA methods in all settings. Our code is available at https://anonymous.4open.science/r/PADA-coder
Deep Learning · Large Language Models
Sparse Autoencoders (SAEs) have become a cornerstone in mechanistic interpretability. However, current training methods inherit the Block Training paradigm from LLM pre-training. We identify this as a critical methodological oversight when applied to instruct models. Theoretically, utilizing GSNR analysis, we prove that attention leakage from unrelated contexts introduces destructive gradient noise. To rectify this paradigm, we propose $\underline{\textbf{F}}$inetuning-$\underline{\textbf{a}}$ligned $\underline{\textbf{S}}$equential $\underline{\textbf{T}}$raining ($\textit{FAST}$), a novel training method specifically tailored for instruct models.$\textit{FAST}$ aligns the training process with the data distribution and activation patterns of instruct models, achieving substantial improvements in both reconstruction and feature interpretability. Experimental results validate the efficacy of$\textit{FAST}$. GSNR analysis confirms improved training performance, with $\textit{FAST}$ demonstrating higher GSNR. This translates into superior reconstruction fidelity: $\textit{FAST}$ achieves an MSE of 0.6468 (significantly outperforming the baseline’s 5.1985) and maintains a near-zero Delta Loss (-0.51% to 0.37%). Consequently, feature quality is markedly enhanced; on Llama-3.2-3B-it, $\textit{FAST}$ yields 21.1% high-quality features, surpassing the 7.0% and 10.2% achieved by baselines. Surprisingly, we discover that intervening on special token activations via SAEs improves output quality, suggesting new opportunities for fine-grained control. Code, data, and all 240 trained SAEs will be publicly released.
What would it mean to have a *scientific* understanding of AI? Language models are not static objects—they are snapshots of time-evolving processes shaped by data, objectives, and optimization dynamics. Yet the field predominantly treats models as fixed artifacts, analyzing behaviors after training rather than asking *why* they emerge. **This position paper argues that AI research should move beyond *post hoc* fixes and study the learning dynamics of models.** We envision a hierarchy of scientific maturity: first *predict* outcomes from early training signals, then *intervene* when trajectories go wrong, ultimately *design* training procedures that guarantee desired properties. Scaling laws have reached the first level for loss; the challenge is extending all three levels to general capabilities, biases, and safety. We articulate requirements for such theories, survey progress across mechanistic interpretability, fairness, memorization, and learning dynamics, and identify concrete open problems. The path forward requires treating models as processes to be understood, not just artifacts to be patched.
Applications · Computer Vision
In most real-world image-to-image (I2I) scenarios, existing evaluations primarily focus on instruction following and the perceptual quality or aesthetics of the generated images. However, they largely fail to assess whether the output image preserves the semantic correspondence and spatial structure of the input image. To address this limitation, we propose StableI2I, a unified and dynamic evaluation framework that explicitly measures content fidelity and pre--post consistency across a wide range of I2I tasks without requiring reference images, including image editing and image restoration. In addition, we construct StableI2I-Bench, a benchmark designed to systematically evaluate the accuracy of MLLMs on such fidelity and consistency assessment tasks. Extensive experimental results demonstrate that StableI2I provides accurate, fine-grained, and interpretable evaluations of content fidelity and consistency, with strong correlations to human subjective judgments. Our framework serves as a practical and reliable evaluation tool for diagnosing content consistency and benchmarking model performance in real-world I2I systems.
Applications · Chemistry, Physics, and Earth Sciences
Inferring cellular trajectories from destructive snapshots is complicated by the challenges of stochasticity and non-conservative mass dynamics such as cell proliferation and apoptosis. Existing unbalanced Optimal Transport (OT) methods treat mass as a continuous fluid, performing inference at the population level. However, this macroscopic view often fails to capture the discrete, jump-like nature of birth-death events at single-cell resolution, which is essential for understanding lineage branching and fate decisions. We present **Unbalanced Schrödinger Bridge (USB)**, a simulation-free framework for learning underlying dynamics that effectively integrates both stochastic and unbalanced effects which also models the discrete, jump-like birth–death dynamics at single-cell resolution. Theoretically, USB provides a tractable solution to the Branching Schrödinger Bridge (BSB) problem, offering a rigorous microscopic interpretation where individual cells undergo both Brownian motion and discrete birth-death jumps. Technically, the method implements an efficient solver by introducing a simulation-free training objective that effectively scales to high-dimensional omics data. Empirically, we demonstrate on both simulated and real-world datasets that USB not only achieves trajectory reconstruction performance better than or comparable to deterministic baselines but also uniquely enables realistic discrete simulation of birth-death dynamics at single-cell resolution.
Applications · Neuroscience, Cognitive Science
Modeling the bidirectional correspondence between external sensory stimuli and internal neural activity has emerged as a critical frontier in neuroscience. However, existing approaches predominantly treat brain encoding and decoding as isolated tasks, relying heavily on unimodal alignment and external priors while overlooking the brain's intrinsic nature as a multimodal integration system. To address these limitations, we propose BrainJanus, the first unified brain foundation model that integrates brain, vision, and language within a single framework. Specifically, we introduce a Unified Brain Tokenizer to quantize continuous neural dynamics into discrete tokens aligned with visual and linguistic representations in a shared Omni space. Building on this, we utilize an All-in-One autoregressive architecture that leverages next-token prediction to enable seamless any-to-any generation, which encompasses image-to-brain and text-to-brain encoding, and brain-to-image and brain-to-text decoding. Extensive experiments demonstrate that BrainJanus achieves superior performance across diverse benchmarks. Furthermore, our framework exhibits zero-shot generalization and preserves interpretable biological topography, highlighting its potential as a general-purpose brain modeling paradigm. The code and pretrained models will be released to support further research.
General Machine Learning · Evaluation
Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns the latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR’s exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.
Deep Learning · Large Language Models
Self-interpretation methods prompt language models to describe their own internal states, but remain unreliable due to hyperparameter sensitivity. We show that training lightweight adapters on interpretability artifacts, while keeping the LM entirely frozen, yields reliable self-interpretation across tasks and model families. A scalar affine adapter with just $d_\text{model}+1$ parameters suffices: trained adapters generate sparse autoencoder feature labels that outperform the training labels themselves (71% vs 63% generation scoring at 70B scale), identify topics with 94% recall@1 versus 1% for untrained baselines, and decode bridge entities in multi-hop reasoning that appear in neither prompt nor response, surfacing implicit reasoning without chain-of-thought. The learned bias vector alone accounts for 85% of improvement, and simpler adapters generalize better than more expressive alternatives. Controlling for model knowledge via prompted descriptions, we find self-interpretation gains outpace capability gains from 7B to 72B parameters. Our results demonstrate that self-interpretation improves with scale, without modifying the model being interpreted.
Social Aspects · Accountability, Transparency, and Interpretability
Activation decomposition methods in language models are tightly coupled to geometric assumptions on how concepts are realized in activation space. Existing approaches search for individual global directions, implicitly assuming linear separability, which overlooks concepts with nonlinear or multi-dimensional structure. In this work, we leverage Mixture of Factor Analyzers (MFA) as a scalable, unsupervised alternative that models the activation space as a collection of Gaussian regions with their *local* covariance structure. MFA decomposes activations into two compositional geometric objects: the region's centroid in activation space, and the local variation from the centroid. We train large-scale MFAs for Llama-3.1-8B and Gemma-2-2B, and show they capture complex, nonlinear structures in activation space. Moreover, evaluations on localization and steering benchmarks show that MFA outperforms unsupervised baselines, is competitive with supervised localization methods, and often achieves stronger steering performance than sparse autoencoders. Together, our findings position local geometry, expressed through subspaces, as a promising unit of analysis for scalable concept discovery and model control, accounting for complex structures that isolated directions fail to capture.
General Machine Learning · Causality
Model Predictive Control (MPC) is widely used to operate safety-critical infrastructure by predicting future trajectories and optimizing control actions. However, nonlinear dynamics, hard safety constraints, and numerical optimization often render individual control moves opaque to human operators, undermining trust and hindering deployment. This paper presents Hierarchical Causal Abduction (HCA), which combines (i) physics-informed reasoning via domain knowledge graphs, (ii) optimization evidence from Karush--Kuhn--Tucker (KKT) multipliers, and (iii) temporal causal discovery via the PCMCI algorithm to generate faithful, human-interpretable explanations for control actions computed by nonlinear MPC. Across three diverse control applications (greenhouse climate, building HVAC, chemical process engineering) with expert validation, HCA improves explanation accuracy by 53\% over LIME (0.478 vs. 0.311) using a single set of cross-domain parameters without per-domain tuning; domain-specific KKT-threshold calibration over 2--3 days further increases accuracy to 0.88. Ablation studies confirm that each evidence source is essential, with 32--37\% accuracy degradation when any component is removed, and HCA's ranking-and-validation methodology generalizes beyond MPC to other prediction-based decision systems, including learning-based control and trajectory planning.
Social Aspects · Accountability, Transparency, and Interpretability
Sparse Autoencoders (SAEs) are widely used to interpret large language models by decomposing activations into sparse, human-understandable features, but scaling to large dictionaries exposes fundamental challenges. Systematic studies reveal pervasive feature splitting that fragments coherent concepts into non-atomic latents and widespread feature absorption that creates arbitrary exceptions in general features, severely compromising latent reliability. These issues stem from inconsistent latent assignment across samples: without cross-sample constraints, per-sample optimization often allows a single underlying concept to be inconsistently distributed across multiple redundant or interfering latents. To address this, we introduce C$^2$R (\underline{\textbf{C}}ross-sample \underline{\textbf{C}}onsistency \underline{\textbf{R}}egularization). C$^2$R explicitly encourages that each semantic feature is consistently represented by a unified latent across the batch by penalizing the co-activation of directionally similar latents. Comprehensive evaluation demonstrates that C$^2$R effectively mitigates both splitting and absorption while, crucially, preserving reconstruction fidelity, providing a principled solution that enhances latent interpretability without degrading model performance. Source code is available\footnote{\url{https://anonymous.4open.science/r/Cross-sample-Consistency-Regularization-28F8}}.
Deep Learning · Large Language Models
Recent advances in Omni-Multimodal Large Language Models (Omni-MLLMs) have enabled strong integration of vision, audio, and language. However, their audio-visual intelligence (AVI) remains insufficiently evaluated due to the lack of systematic and comprehensive benchmarks. We introduce AVI-Bench, a cognitively inspired benchmark that evaluates Omni-MLLMs across three stages, perception, understanding, and reasoning, through cross-modal tasks requiring joint audio-visual interpretation. This design enables fine-grained diagnosis of model capabilities and failure modes. To further assess robustness beyond familiar domains, we propose AVI-Bench-PriSe, an extension that probes models' primitive audio-visual sensation using unfamiliar, low-semantic stimuli, testing generalization beyond common training distributions. Extensive experiments on both open-source and closed-source models reveal substantial limitations in current Omni-MLLMs. Based on these findings, we present a four-level AVI taxonomy. Overall, AVI-Bench provides a principled evaluation framework to guide the development of more robust and generalizable AVI.
Deep Learning · Foundation Models
Existing genomic foundation models (GFMs) typically treat DNA as raw nucleotide sequences, often overlooking the regulatory context required to interpret genetic variation accurately. We introduce BioToken, a tokenization framework that directly encodes variants and biological annotations into genomic representations, and BioFM, a parameter-efficient model built on this architecture. By leveraging biological inductive biases, BioFM outperforms state-of-the-art models and specialized baselines like Enformer on benchmarks including pathogenicity and expression prediction while requiring 100-fold less compute than current large-scale genomic models. These findings demonstrate that explicitly modeling biological structure yields more robust and efficient genomic representations than scaling alone.
Social Aspects · Accountability, Transparency, and Interpretability
Many visual explanation methods in computer vision highlight pixel importance but struggle to link these low-level cues to semantically meaningful concepts, limiting their interpretability and trustworthiness. We introduce Concept-based Explanations (ConEx), a novel framework that bridges saliency visualization with concept-based reasoning to provide both localized and global interpretability. ConEx automatically discovers class-specific concepts and represents them through concept activation vectors (CAVs), learned without manual supervision using an architecture-specific masking mechanism that reduces noise introduced by the segmentation masks to enhance concept purity. Locally, ConEx generates saliency maps that reveal where each concept appears in the image and how it contributes to the prediction; globally, it identifies the most influential concepts for each class. To evaluate the reliability of these learned concepts, we propose two complementary metrics, Vector-Concept Match (VCM) and Concept-Class Match (CCM), that quantify concept alignment and enable direct comparison with existing methods. Extensive experiments across diverse datasets and architectures demonstrate that ConEx achieves state-of-the-art performance on faithfulness, segmentation, and concept-quality benchmarks. Human studies further confirm that the discovered concepts are interpretable, distinctive, and aligned with human understanding. Overall, ConEx advances the field toward truly interpretable and concept-grounded explanations in vision models.
Social Aspects · Accountability, Transparency, and Interpretability
While *Prover-Verifier Games* (PVGs) offer a promising path toward verifiability in nonlinear classification models, they have not yet been applied to complex inputs such as high-dimensional images. Conversely, expressive *concept encodings* effectively allow to translate such data into interpretable concepts but are often utilised in the context of low-capacity linear predictors. In this work, we push towards real-world verifiability by combining the strengths of both approaches. We introduce *Neural Concept Verifier (NCV)*, a unified framework combining PVGs for formal verifiability with concept encodings to handle complex, high-dimensional inputs in an interpretable way. NCV achieves this by utilizing recent minimally supervised concept discovery models to extract structured concept encodings from raw inputs. A *prover* then selects a subset of these encodings, which a *verifier*, implemented as a nonlinear predictor, uses exclusively for decision-making. Our evaluations show that NCV outperforms classic concept-based models and pixel-based PVG classifier baselines on high-dimensional, logically complex datasets and helps mitigate shortcut behavior. Overall, we demonstrate NCV as a promising step toward concept-level, verifiable AI.
Deep Learning · Large Language Models
In modern software development, particularly in emerging ``vibe coding'' paradigms, project implementation increasingly begins with visual interactions between users and AI coding assistants, where system architectures are communicated through visual designs before coding. This visual-first approach necessitates AI systems capable of interpreting diagrams across multiple programming languages. However, the development of such systems is severely hindered by the lack of large-scale multimodal training data and evaluation benchmarks. To address these limitations, we present M2C-INSTRUCT, a comprehensive multilingual multimodal instruction-tuning dataset containing over 13.1M samples across 50+ programming languages, designed for visual understanding and diagram interpretation in code generation tasks. We validate our dataset by training M2-CODER, a multilingual multimodal software developer that successfully integrates visual design inputs with textual instructions. We also introduce M2EVAL, a novel multilingual evaluation benchmark for multimodal code generation performance. Experiments show our 7B M2-CODER performs on par with much larger 70B+ models, confirming the quality and effectiveness of our M2C-INSTRUCT. Together, M2C-INSTRUCT, M2-CODER, and M2EVAL provide essential infrastructure for visual-assisted programming in vibe-coding and visual-interactive development workflows.
Deep Learning · Large Language Models
The superficial alignment hypothesis (SAH) posits that large language models learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge. The SAH, however, lacks a precise definition, which has led to (i) different and seemingly orthogonal arguments supporting it, and (ii) important critiques to it. We propose a new metric called **Task Complexity**: the length of the shortest program that achieves a target performance on a task. In this framework, the SAH claims that pre-trained models drastically reduce the task complexity of achieving high performance on many tasks. Our definition unifies prior arguments supporting the SAH, interpreting them as different strategies to find such short programs. Experimentally, we estimate task complexities of mathematical reasoning, machine translation, and instruction following tasks and show that their respective task complexities can be remarkably low when conditioned on a pre-trained model. Further, we find that pre-training enables access to strong performances on our tasks, but it can require programs of gigabytes of length to access them. Post-training, on the other hand, collapses the complexity of reaching this same performance by several orders of magnitude. Overall, our results highlight that task adaptation can require remarkably little information—often just a few kilobytes.
Deep Learning · Generative Models and Autoencoders
Text-to-image (T2I) diffusion models (DMs) have achieved remarkable generative quality but still exhibit the risk to produce biased and inappropriate images. A promising line of prior work aims to mitigate this issue by learning interpretable and linearly controllable concepts from semantic spaces, such as the U-Net bottleneck. However, these methods rely entirely on the bottleneck layer in U-Net and therefore cannot be directly applied to modern ViT-based DMs, including FLUX and PixArt. In this work, we present a model agnostic framework for discovering interpretable and linearly controllable semantic attributes across any T2I DMs backbone. We first show that multi-modal attention heads in ViT-based DMs encode interpretable and (near) linear semantic structures similar to those in the U-Net bottleneck. Based on this insight, we introduce a method that learns external concept vectors, which are added to the multi-modal attention heads for ViT-based DMs or to the bottleneck layer for U-Net-based DMs, while keeping the pretrained model frozen. Experiments across SDXL, SD3.5, PixArt, and FLUX demonstrate that these concept vectors provide interpretability, linearity, and highly improved fairness while preserving visual fidelity. The code is included in the supplementary material.
General Machine Learning · Evaluation
Evaluation benchmarks play a central role in assessing vision–language models (VLMs). However, most existing multimodal benchmarks are static, making them increasingly vulnerable to data contamination, temporal staleness, and high construction costs. In this work, we introduce MMBench-Live, a multi-agent-driven dynamic multimodal benchmark that supports continuous updates without human in the loop. MMBench-Live is maintained through an end-to-end automated pipeline that integrates structured benchmark description, real-time data acquisition, and verifiable question–answer (QA) generation, enabling scalable, live, and low-cost benchmark updates. To ensure reliable evaluation across versions, we further propose a distribution-consistent updating strategy based on semantic task interpretation and feedback-driven data collection and filtering. Based on MMBench-Live, we conduct systematic evaluations of multiple open-source VLMs and analyze their performance, cross-version consistency, and data contamination, providing empirical evidence for the effectiveness of the proposed dynamic benchmark updating framework.