Your paper timeline
Scroll AI takes the way you would scroll a great paper aggregator: quick signal first, deeper critique when something earns your attention, and challenges when a claim feels off.
181 papers in cs.LG
Trending mixes fresh papers with community signal.
0
cs.LGcs.AI Kexin Huang, Haoming Meng, Junkang Wu et al. · Mar 23, 2026

This paper investigates how Reinforcement Learning with Verifiable Rewards (RLVR) improves LLM reasoning by focusing on the *direction* of policy updates rather than their magnitude. The authors introduce $\Delta \log p$, the signed log-probability difference between base and RLVR models, and argue it better captures reasoning-critical tokens than magnitude-based metrics like entropy or KL divergence. They validate this through token-replacement interventions and propose two practical applications: a test-time extrapolation method that amplifies the learned direction without additional training, and a training-time reweighting scheme that focuses learning on low-probability tokens.

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models. While existing analyses identify that RLVR-induced changes are sparse, they primarily focus on the \textbf{magnitude} of these updates, largely overlooking their \textbf{direction}. In this work, we argue that the direction of updates is a more critical lens for understanding RLVR's effects, which can be captured by the signed, token-level log probability difference $\Delta\log p$ between the base and final RLVR models. Through statistical analysis and token-replacement interventions, we demonstrate that $\Delta\log p$ more effectively identifies sparse, yet reasoning-critical updates than magnitude-based metrics (\eg divergence or entropy). Building on this insight, we propose two practical applications: (1) a \textit{test-time extrapolation} method that amplifies the policy along the learned $\Delta\log p$ direction to improve reasoning accuracy without further training; (2) a \textit{training-time reweighting} method that focuses learning on low-probability (corresponding to higher $\Delta\log p$) tokens, which improves reasoning performance across models and benchmarks. Our work establishes the direction of change as a key principle for analyzing and improving RLVR.
0
cs.LGcs.AIcs.CL Kexian Tang, Jiani Wang, Shaowen Wang et al. · Mar 23, 2026

Large language models often lack coverage in specialized, data-scarce domains where web text is limited. This paper proposes SPA (Scaling Prompt-engineered Augmentation), a baseline that generates large-scale synthetic corpora using just seven carefully designed prompt templates grounded in cognitive learning strategies (Concept Learning, Critical Thinking, and Generative Learning). The core finding is that this simple approach consistently outperforms complex RL-based methods like SEAL and multi-stage pipelines like EntiGraph across Wikipedia QA, long-document comprehension, and multi-hop reasoning benchmarks, suggesting that careful prompt design combined with straightforward scaling is surprisingly effective for knowledge injection.

While large language models (LLMs) are pretrained on massive amounts of data, their knowledge coverage remains incomplete in specialized, data-scarce domains, motivating extensive efforts to study synthetic data generation for knowledge injection. We propose SPA (Scaling Prompt-engineered Augmentation), a simple but tough-to-beat baseline that uses a small set of carefully designed prompts to generate large-scale synthetic data for knowledge injection. Through systematic comparisons, we find that SPA outperforms several strong baselines. Furthermore, we identify two key limitations of prior approaches: (1) while RL-based methods may improve the token efficiency of LLM-based data augmentation at small scale, they suffer from diversity collapse as data scales, leading to diminishing returns; and (2) while multi-stage prompting may outperform simple augmentation methods, their advantages can disappear after careful prompt tuning. Our results suggest that, for knowledge injection, careful prompt design combined with straightforward large-scale augmentation can be surprisingly effective, and we hope SPA can serve as a strong baseline for future studies in this area. Our code is available at https://github.com/Tangkexian/SPA.
0
cs.CVcs.AIcs.LG Donald Shenaj, Federico Errica, Antonio Carta · Mar 23, 2026

Personalized image generation with diffusion models relies on Low-Rank Adaptation (LoRA) to fine-tune models efficiently, but current practice uses a fixed rank across all layers regardless of subject complexity. This paper proposes LoRA2, which learns adaptive ranks per LoRA component via a variational framework that imposes an importance ordering over rank indices using a discretized exponential distribution. The method achieves better subject fidelity and prompt alignment while using significantly less memory than high-rank baselines, addressing the combinatorial explosion of searching $S K^L$ architectural configurations.

Low Rank Adaptation (LoRA) is the de facto fine-tuning strategy to generate personalized images from pre-trained diffusion models. Choosing a good rank is extremely critical, since it trades off performance and memory consumption, but today the decision is often left to the community's consensus, regardless of the personalized subject's complexity. The reason is evident: the cost of selecting a good rank for each LoRA component is combinatorial, so we opt for practical shortcuts such as fixing the same rank for all components. In this paper, we take a first step to overcome this challenge. Inspired by variational methods that learn an adaptive width of neural networks, we let the ranks of each layer freely adapt during fine-tuning on a subject. We achieve it by imposing an ordering of importance on the rank's positions, effectively encouraging the creation of higher ranks when strictly needed. Qualitatively and quantitatively, our approach, LoRA$^2$, achieves a competitive trade-off between DINO, CLIP-I, and CLIP-T across 29 subjects while requiring much less memory and lower rank than high rank LoRA versions. Code: https://github.com/donaldssh/NotAllLayersAreCreatedEqual.
0
cs.LGcs.AI Dilina Rajapakse, Juan C. Rosero, Ivana Dusparic · Mar 23, 2026

Multi-Objective Reinforcement Learning (MORL) agents must balance competing objectives like speed versus energy consumption, yet existing Explainable RL methods fail to clarify how specific behavioral choices drive Pareto trade-offs. This paper proposes TREX, a post-hoc trajectory attribution framework that clusters agent behaviors into semantically meaningful segments and quantifies each cluster's influence on objective trade-offs by training complementary policies that exclude specific trajectory groups. The work addresses a genuine gap in explainability by moving beyond policy selection to reveal which behavioral patterns (such as "long leaps" versus "short strides") justify the agent's learned trade-off logic.

Reinforcement Learning (RL) has demonstrated its ability to solve complex decision-making problems in a variety of domains, by optimizing reward signals obtained through interaction with an environment. However, many real-world scenarios involve multiple, potentially conflicting objectives that cannot be easily represented by a single scalar reward. Multi-Objective Reinforcement Learning (MORL) addresses this limitation by enabling agents to optimize several objectives simultaneously, explicitly reasoning about trade-offs between them. However, the ``black box" nature of the RL models makes the decision process behind chosen objective trade-offs unclear. Current Explainable Reinforcement Learning (XRL) methods are typically designed for single scalar rewards and do not account for explanations with respect to distinct objectives or user preferences. To address this gap, in this paper we propose TREX, a Trajectory based Explainability framework to explain Multi-objective Reinforcement Learning policies, based on trajectory attribution. TREX generates trajectories directly from the learned expert policy, across different user preferences and clusters them into semantically meaningful temporal segments. We quantify the influence of these behavioural segments on the Pareto trade-off by training complementary policies that exclude specific clusters, measuring the resulting relative deviation on the observed rewards and actions compared to the original expert policy. Experiments on multi-objective MuJoCo environments - HalfCheetah, Ant and Swimmer, demonstrate the framework's ability to isolate and quantify the specific behavioural patterns.
0
cs.LGcs.AIcs.CV Bahar Dibaei Nia, Farzan Farnia · Mar 23, 2026

The paper addresses sample-efficient selection among multiple pretrained generative models, formulated as a diversity-aware multi-armed bandit problem where the optimal solution may be a mixture rather than a single model. The authors challenge the necessity of explicit UCB exploration bonuses, proposing that Mixture-Greedy—which directly optimizes empirical diversity objectives without optimism bonuses—can achieve sublinear regret through implicit exploration induced by the objective geometry. This matters because sampling from suboptimal generative models is computationally expensive, and their results suggest that structural properties of diversity metrics (FID, Vendi, RKE) naturally enforce sufficient exploration without costly confidence bound computations.

Efficient selection among multiple generative models is increasingly important in modern generative AI, where sampling from suboptimal models is costly. This problem can be formulated as a multi-armed bandit task. Under diversity-aware evaluation metrics, a non-degenerate mixture of generators can outperform any individual model, distinguishing this setting from classical best-arm identification. Prior approaches therefore incorporate an Upper Confidence Bound (UCB) exploration bonus into the mixture objective. However, across multiple datasets and evaluation metrics, we observe that the UCB term consistently slows convergence and often reduces sample efficiency. In contrast, a simple \emph{Mixture-Greedy} strategy without explicit UCB-type optimism converges faster and achieves even better performance, particularly for widely used metrics such as FID and Vendi where tight confidence bounds are difficult to construct. We provide theoretical insight explaining this behavior: under transparent structural conditions, diversity-aware objectives induce implicit exploration by favoring interior mixtures, leading to linear sampling of all arms and sublinear regret guarantees for entropy-based, kernel-based, and FID-type objectives. These results suggest that in diversity-aware multi-armed bandits for generative model selection, exploration can arise intrinsically from the objective geometry, questioning the necessity of explicit confidence bonuses.
0
cs.LGcs.AIstat.ME Marc Franquesa Mon\'es, Jiaqi Zhang, Caroline Uhler · Mar 23, 2026

Constraint-based causal discovery algorithms like PC require exponentially many conditional independence (CI) tests in the worst case---specifically $p^{\mathcal{O}(d)}$ where $d$ is the maximum degree. This paper establishes that the fundamental complexity parameter is actually $s$, the maximum undirected clique size in the essential graph, which can be much smaller than $d$ (e.g., $s=2$ vs $d=p-2$ in Figure 1). The authors propose Greedy Ancestral Search (GAS), which achieves $p^{\mathcal{O}(s)}$ CI tests, and prove a matching lower bound of $2^{\Omega(s)}$, establishing exponent-optimality up to a logarithmic factor.

Learning causal relations from observational data is a fundamental problem with wide-ranging applications across many fields. Constraint-based methods infer the underlying causal structure by performing conditional independence tests. However, existing algorithms such as the prominent PC algorithm need to perform a large number of independence tests, which in the worst case is exponential in the maximum degree of the causal graph. Despite extensive research, it remains unclear if there exist algorithms with better complexity without additional assumptions. Here, we establish an algorithm that achieves a better complexity of $p^{\mathcal{O}(s)}$ tests, where $p$ is the number of nodes in the graph and $s$ denotes the maximum undirected clique size of the underlying essential graph. Complementing this result, we prove that any constraint-based algorithm must perform at least $2^{\Omega(s)}$ conditional independence tests, establishing that our proposed algorithm achieves exponent-optimality up to a logarithmic factor in terms of the number of conditional independence tests needed. Finally, we validate our theoretical findings through simulations, on semi-synthetic gene-expression data, and real-world data, demonstrating the efficiency of our algorithm compared to existing methods in terms of number of conditional independence tests needed.
0
cs.LGcs.AIcs.CL Hung-Hsuan Chen · Mar 23, 2026

Standard Transformers apply fixed-depth computation regardless of problem difficulty, limiting their ability to solve tasks requiring variable-depth reasoning like multi-hop traversal or nested logic. This paper proposes a depth-recurrent Transformer that iteratively applies a shared-weight block in latent space—enabling 'vertical Chain-of-Thought' where models trade recurrence steps for deeper reasoning without consuming context window. The work demonstrates strong compositional generalization on three synthetic tasks and offers a mechanistic alternative to horizontal token-generation paradigms.

Standard Transformers have a fixed computational depth, fundamentally limiting their ability to generalize to tasks requiring variable-depth reasoning, such as multi-hop graph traversal or nested logic. We propose a depth-recurrent Transformer that decouples computational depth from parameter count by iteratively applying a shared-weight Transformer block in latent space -- enabling the model to trade recurrence steps for deeper reasoning at inference time. Our architecture incorporates three mechanisms to make deep recurrence (20+ steps) stable: (1) a silent thinking objective that supervises only the final output, forcing genuine multi-step reasoning rather than intermediate heuristic shortcuts; (2) LayerScale initialization to protect fragile reasoning states from untrained layer noise; and (3) an identity-biased recurrence that creates a gradient highway across many steps. We evaluate on three compositional reasoning domains with decreasing inductive biases: graph reachability (strict adjacency masking), nested boolean logic (relative positioning), and unstructured relational text (where sequence position provides no structural hints). Across all tasks, we observe a clear \emph{computational frontier} -- a boundary where performance transitions from chance to near-perfect as thinking steps scale with task complexity. Moreover, these tasks reveal qualitatively different generalization behaviors: precise but brittle (graph), approximate but robust (logic), and autonomous latent routing without structural hints (text). This progression illuminates how the interplay between a task-invariant recurrent reasoning core and task-specific perceptual interfaces shapes out-of-distribution (OOD) generalization, offering a mechanistic perspective on vertical chain-of-thought that complements the prevailing horizontal token-generation paradigm.
0
physics.opticscs.AIcs.AR Hyoseok Park, Yeonsang Park · Mar 23, 2026

Long-context LLM inference hits a memory wall: each decode step requires scanning the entire KV cache, incurring $O(n)$ memory bandwidth that cannot be solved by faster arithmetic. PRISM proposes a thin-film lithium niobate photonic accelerator that performs the block-selection similarity search in $O(1)$ optical latency using a broadcast-and-weight architecture, eliminating the $O(n)$ scan entirely. The work claims $16\times$–$32\times$ traffic reduction at 64K–128K tokens and a four-order-of-magnitude energy advantage over GPU baselines by matching photonic hardware capabilities—passive query broadcast, quasi-static microring weights, and low-precision rank output—to the selection task.

Long-context LLM inference is bottlenecked not by compute but by the O(n) memory bandwidth cost of scanning the KV cache at every decode step -- a wall that no amount of arithmetic scaling can break. Recent photonic accelerators have demonstrated impressive throughput for dense attention computation; however, these approaches inherit the same O(n) memory scaling as electronic attention when applied to long contexts. We observe that the real leverage point is the coarse block-selection step: a memory-bound similarity search that determines which KV blocks to fetch. We identify, for the first time, that this task is structurally matched to the photonic broadcast-and-weight paradigm -- the query fans out to all candidates via passive splitting, signatures are quasi-static (matching electro-optic MRR programming), and only rank order matters (relaxing precision to 4-6 bits). Crucially, the photonic advantage grows with context length: as N increases, the electronic scan cost rises linearly while the photonic evaluation remains O(1). We instantiate this insight in PRISM (Photonic Ranking via Inner-product Similarity with Microring weights), a thin-film lithium niobate (TFLN) similarity engine. Hardware-impaired needle-in-a-haystack evaluation on Qwen2.5-7B confirms 100% accuracy from 4K through 64K tokens at k=32, with 16x traffic reduction at 64K context. PRISM achieves a four-order-of-magnitude energy advantage over GPU baselines at practical context lengths (n >= 4K).
0
cs.CVcs.AIcs.GR Kangbo Zhao, Miaoxin Guan, Xiang Chen et al. · Mar 23, 2026

This paper tackles the domain generalization problem in image deraining, where models trained on synthetic data fail catastrophically on out-of-distribution (OOD) real-world scenarios. The authors propose a three-stage pipeline—Superpixel Generation, Resolution-adaptive Fusion, and Pseudo-label Re-Synthesis—that adapts source-domain models to target domains using only unpaired rain-free images, eliminating the need for costly paired rainy data collection.

Image deraining plays a pivotal role in low-level computer vision, serving as a prerequisite for robust outdoor surveillance and autonomous driving systems. While deep learning paradigms have achieved remarkable success in firmly aligned settings, they often suffer from severe performance degradation when generalized to unseen Out-of-Distribution (OOD) scenarios. This failure stems primarily from the significant domain discrepancy between synthetic training datasets and the complex physical dynamics of real-world rain. To address these challenges, this paper proposes a pioneering cross-scenario deraining adaptation framework. Diverging from conventional approaches, our method obviates the requirements for paired rainy observations in the target domain, leveraging exclusively rain-free background images. We design a Superpixel Generation (Sup-Gen) module to extract stable structural priors from the source domain using Simple Linear Iterative Clustering. Subsequently, a Resolution-adaptive Fusion strategy is introduced to align these source structures with target backgrounds through texture similarity, ensuring the synthesis of diverse and realistic pseudo-data. Finally, we implement a pseudo-label re-Synthesize mechanism that employs multi-stage noise generation to simulate realistic rain streaks. This framework functions as a versatile plug-and-play module capable of seamless integration into arbitrary deraining architectures. Extensive experiments on state-of-the-art models demonstrate that our approach yields remarkable PSNR gains of up to 32% to 59% in OOD domains while significantly accelerating training convergence.
0
cs.LGcs.AI Yawen Li, Tao Hu, Zhouhui Lian et al. · Mar 23, 2026

This paper studies generalization error bounds for Transformer models using offset Rademacher complexity. The core idea is to derive sharp excess risk bounds that achieve optimal $O(1/n)$ convergence rates---improving upon the standard $O(1/\sqrt{n})$---for single-head, multi-head, and multi-layer architectures, with explicit dependence on matrix ranks and parameter norms. The authors further extend these results to unbounded sub-Gaussian and heavy-tailed input distributions, broadening the applicability beyond standard boundedness assumptions.

This paper studies generalization error bounds for Transformer models. Based on the offset Rademacher complexity, we derive sharper generalization bounds for different Transformer architectures, including single-layer single-head, single-layer multi-head, and multi-layer Transformers. We first express the excess risk of Transformers in terms of the offset Rademacher complexity. By exploiting its connection with the empirical covering numbers of the corresponding hypothesis spaces, we obtain excess risk bounds that achieve optimal convergence rates up to constant factors. We then derive refined excess risk bounds by upper bounding the covering numbers of Transformer hypothesis spaces using matrix ranks and matrix norms, leading to precise, architecture-dependent generalization bounds. Finally, we relax the boundedness assumption on feature mappings and extend our theoretical results to settings with unbounded (sub-Gaussian) features and heavy-tailed distributions.
0
cs.LGcs.AIstat.ML Abdou-Raouf Atarmla · Mar 23, 2026

Rule-State Inference (RSI) addresses compliance monitoring in domains like taxation where authoritative rules are known a priori but observations are partial, noisy, or strategically distorted. The paper proposes a Bayesian framework that inverts the standard ML paradigm: instead of learning rules from data, RSI encodes regulatory rules as structured priors and infers latent rule states (activation, compliance rate, parametric drift) via posterior inference. This enables zero-shot compliance assessment without labeled training data—a critical capability for low-resource environments where non-compliance labels are scarce or legally sensitive.

Existing machine learning frameworks for compliance monitoring -- Markov Logic Networks, Probabilistic Soft Logic, supervised models -- share a fundamental paradigm: they treat observed data as ground truth and attempt to approximate rules from it. This assumption breaks down in rule-governed domains such as taxation or regulatory compliance, where authoritative rules are known a priori and the true challenge is to infer the latent state of rule activation, compliance, and parametric drift from partial and noisy observations. We propose Rule-State Inference (RSI), a Bayesian framework that inverts this paradigm by encoding regulatory rules as structured priors and casting compliance monitoring as posterior inference over a latent rule-state space S = {(a_i, c_i, delta_i)}, where a_i captures rule activation, c_i models the compliance rate, and delta_i quantifies parametric drift. We prove three theoretical guarantees: (T1) RSI absorbs regulatory changes in O(1) time via a prior ratio correction, independently of dataset size; (T2) the posterior is Bernstein-von Mises consistent, converging to the true rule state as observations accumulate; (T3) mean-field variational inference monotonically maximizes the Evidence Lower BOund (ELBO). We instantiate RSI on the Togolese fiscal system and introduce RSI-Togo-Fiscal-Synthetic v1.0, a benchmark of 2,000 synthetic enterprises grounded in real OTR regulatory rules (2022-2025). Without any labeled training data, RSI achieves F1=0.519 and AUC=0.599, while absorbing regulatory changes in under 1ms versus 683-1082ms for full model retraining -- at least a 600x speedup.
0
cs.LGcs.AI Bayezid Baten, M. Ayyan Iqbal, Sebastian Ament et al. · Mar 23, 2026

Concrete mix design requires balancing competing objectives of mechanical strength and sustainability. BOxCrete introduces a Gaussian Process regression framework trained on 533 strength measurements from 123 unique mixtures to predict compressive strength evolution over curing time and optimize mixes for embodied carbon using multi-objective Bayesian Optimization. The work addresses a critical gap in the literature by providing an open-source alternative to proprietary industrial datasets and models.

Modern concrete must simultaneously satisfy evolving demands for mechanical performance, workability, durability, and sustainability, making mix designs increasingly complex. Recent studies leveraging Artificial Intelligence (AI) and Machine Learning (ML) models show promise for predicting compressive strength and guiding mix optimization, but most existing efforts are based on proprietary industrial datasets and closed-source implementations. Here we introduce BOxCrete, an open-source probabilistic modeling and optimization framework trained on a new open-access dataset of over 500 strength measurements (1-15 ksi) from 123 mixtures - 69 mortar and 54 concrete mixes tested at five curing ages (1, 3, 5, 14, and 28 days). BOxCrete leverages Gaussian Process (GP) regression to predict strength development, achieving average R$^2$ = 0.94 and RMSE = 0.69 ksi, quantify uncertainty, and carry out multi-objective optimization of compressive strength and embodied carbon. The dataset and model establish a reproducible open-source foundation for data-driven development of AI-based optimized mix designs.
0
cs.CLcs.AIcs.LG Mariela M. Nina, Caio Veloso Costa, Lilian Berton et al. · Mar 22, 2026

This paper addresses computational barriers for Brazilian Portuguese question answering by systematically evaluating Parameter-Efficient Fine-Tuning (PEFT) methods on BERTimbau models using the SQuAD-BR dataset. The authors test LoRA, DoRA, QLoRA, and QDoRA across Base (110M) and Large (335M) variants, demonstrating that LoRA achieves 95.8% of full fine-tuning performance while reducing training time by 73.5%. A key finding is that PEFT methods require substantially higher learning rates ($2\times 10^{-4}$) than standard BERT fine-tuning to achieve optimal results, with quantization resilience favoring larger models.

Although large language models have transformed natural language processing, their computational costs create accessibility barriers for low-resource languages such as Brazilian Portuguese. This work presents a systematic evaluation of Parameter-Efficient Fine-Tuning (PEFT) and quantization techniques applied to BERTimbau for Question Answering on SQuAD-BR, the Brazilian Portuguese translation of SQuAD v1. We evaluate 40 configurations combining four PEFT methods (LoRA, DoRA, QLoRA, QDoRA) across two model sizes (Base: 110M, Large: 335M parameters). Our findings reveal three critical insights: (1) LoRA achieves 95.8\% of baseline performance on BERTimbau-Large while reducing training time by 73.5\% (F1=81.32 vs 84.86); (2) higher learning rates (2e-4) substantially improve PEFT performance, with F1 gains of up to +19.71 points over standard rates; and (3) larger models show twice the quantization resilience (loss of 4.83 vs 9.56 F1 points). These results demonstrate that encoder-based models can be efficiently fine-tuned for extractive Brazilian Portuguese QA with substantially lower computational cost than large generative LLMs, promoting more sustainable approaches aligned with \textit{Green AI} principles. An exploratory evaluation of Tucano and Sabi\'a on the same extractive QA benchmark shows that while generative models can reach competitive F1 scores with LoRA fine-tuning, they require up to 4.2$\times$ more GPU memory and 3$\times$ more training time than BERTimbau-Base, reinforcing the efficiency advantage of smaller encoder-based architectures for this task.
0
cs.CVcs.AIcs.DB Mohammad Eslami, Dhanvinkumar Ganeshkumar, Saber Kazeminasab et al. · Mar 23, 2026

CataractSAM-2 adapts Meta's Segment Anything Model 2 (SAM-2) for real-time semantic segmentation in cataract surgery videos. The core idea is to fine-tune only the prompt encoder and mask decoder while freezing the image encoder, enabling precise segmentation of anatomical structures and surgical instruments under challenging conditions like glare and occlusion. The paper also introduces an interactive annotation framework that propagates sparse user prompts across video frames to accelerate ground-truth generation.

We present CataractSAM-2, a domain-adapted extension of Meta's Segment Anything Model 2, designed for real-time semantic segmentation of cataract ophthalmic surgery videos with high accuracy. Positioned at the intersection of computer vision and medical robotics, CataractSAM-2 enables precise intraoperative perception crucial for robotic-assisted and computer-guided surgical systems. Furthermore, to alleviate the burden of manual labeling, we introduce an interactive annotation framework that combines sparse prompts with video-based mask propagation. This tool significantly reduces annotation time and facilitates the scalable creation of high-quality ground-truth masks, accelerating dataset development for ocular anterior segment surgeries. We also demonstrate the model's strong zero-shot generalization to glaucoma trabeculectomy procedures, confirming its cross-procedural utility and potential for broader surgical applications. The trained model and annotation toolkit are released as open-source resources, establishing CataractSAM-2 as a foundation for expanding anterior ophthalmic surgical datasets and advancing real-time AI-driven solutions in medical robotics, as well as surgical video understanding.
0
cs.LGcs.AI Woosung Koh, Jeyoung Jeon, Youngjin Song et al. · Mar 23, 2026

The paper tackles the inefficiency of homogeneous compute allocation in multi-task supervised fine-tuning (SFT), where fast-learning tasks overfit while slow ones remain under-trained. The authors propose mSFT, an iterative algorithm that dynamically excludes overfitting sub-datasets and reverts to optimal checkpoints. Their approach consistently outperforms baselines across 6 models and 10 benchmarks, sometimes reducing compute while improving accuracy.

Current language model training commonly applies multi-task Supervised Fine-Tuning (SFT) using a homogeneous compute budget across all sub-datasets. This approach is fundamentally sub-optimal: heterogeneous learning dynamics cause faster-learning tasks to overfit early while slower ones remain under-fitted. To address this, we introduce mSFT, an iterative, overfitting-aware search algorithm for multi-task data mixtures. mSFT trains the model on an active mixture, identifies and excludes the earliest overfitting sub-dataset, and reverts to that specific optimal checkpoint before continuing. Extensive evaluations demonstrate that mSFT consistently outperforms 4 baselines across 10 benchmarks and 6 base models. Further analysis confirms mSFT maintains robust gains across diverse dataset sizes, task granularities, and is insensitive to its single new hyperparameter (compute budget). Notably, at low compute budget, mSFT can improve performance while lowering training FLOPs. Ultimately, mSFT establishes a practical overfitting-aware algorithm for multi-task SFT that maximizes the potential of models across diverse data mixtures.
0
cs.AIcs.CLcs.LG Wihan van der Heever, Keane Ong, Ranjan Satapathy et al. · Mar 23, 2026

This paper addresses the fundamental problem that correlational sentiment analysis cannot distinguish genuine economic associations from spurious statistical artifacts in financial markets. The core contribution is a refutation-validated framework for aspect-based sentiment analysis that combines net-ratio sentiment scoring with four robustness tests—placebo, random common cause, subset stability, and bootstrap validation—to filter false discoveries in high-dimensional sentiment-return analysis. This matters because investment strategies built on spurious correlations can lead to systematic losses, and regulators increasingly demand explainable AI systems with auditable validation.

This paper proposes a refutation-validated framework for aspect-based sentiment analysis in financial markets, addressing the limitations of correlational studies that cannot distinguish genuine associations from spurious ones. Using X data for the energy sector, we test whether aspect-level sentiment signals show robust, refutation-validated relationships with equity returns. Our pipeline combines net-ratio scoring with z-normalization, OLS with Newey West HAC errors, and refutation tests including placebo, random common cause, subset stability, and bootstrap. Across six energy tickers, only a few associations survive all checks, while renewables show aspect and horizon specific responses. While not establishing causality, the framework provides statistically robust, directionally interpretable signals, with limited sample size (six stocks, one quarter) constraining generalizability and framing this work as a methodological proof of concept.
0
cs.AIcs.CRcs.LG Gregory M. Ruddell · Mar 22, 2026

This paper introduces "silent commitment failure" — a phenomenon where instruction-tuned language models produce confident, incorrect outputs with no detectable pre-commitment warning signal — and proposes "governability" as a measurable property for AI agent safety. The core claim is that 2 of 3 instruction-following models evaluated exhibit zero-warning failure modes, with profound implications for autonomous agent deployment. The work distinguishes itself from hallucination studies by focusing on detectability before commitment rather than correctness of output, and presents empirical evidence that conflict-detection signals (the "authority band") are geometric properties fixed at pretraining rather than injectable through fine-tuning.

As large language models are deployed as autonomous agents with tool execution privileges, a critical assumption underpins their security architecture: that model errors are detectable at runtime. We present empirical evidence that this assumption fails for two of three instruction-following models evaluable for conflict detection. We introduce governability -- the degree to which a model's errors are detectable before output commitment and correctable once detected -- and demonstrate it varies dramatically across models. In six models across twelve reasoning domains, two of three instruction-following models exhibited silent commitment failure: confident, fluent, incorrect output with zero warning signal. The remaining model produced a detectable conflict signal 57 tokens before commitment under greedy decoding. We show benchmark accuracy does not predict governability, correction capacity varies independently of detection, and identical governance scaffolds produce opposite effects across models. A 2x2 experiment shows a 52x difference in spike ratio between architectures but only +/-0.32x variation from fine-tuning, suggesting governability is fixed at pretraining. We propose a Detection and Correction Matrix classifying model-task combinations into four regimes: Governable, Monitor Only, Steer Blind, and Ungovernable.
0
cs.LGcs.AI Philip S. Yu, Li Sun · Mar 23, 2026

This paper proposes Riemannian Foundation Model (RFM), a vision for unifying graph learning through Riemannian geometry rather than GNN message-passing or LLM serialization. The authors argue that graphs are discrete analogs of manifolds, and that concepts like vector bundles, curvature, and parallel transport provide the proper toolkit for universal graph modeling—enabling both structural inference and generation in a way that current Euclidean GNNs and tokenized LLMs cannot achieve.

Graphs provide a natural description of the complex relationships among objects, and play a pivotal role in communications, transportation, social computing, the life sciences, etc. Currently, there is strong agreement that Graph Foundation Models (GFMs) are essential for advancing graph learning, yet considerable disagreement persists on how to build a powerful, general-purpose GFM analogous to Large Language Models (LLMs). Graph Neural Networks (GNNs) exhibit limitations in memory retention and principled interpretability when confronted with multi-domain pretraining and adaptation. The challenge of graph serialization hinders the direct application of LLMs, as the words struggle to capture the structural complexity and diversity inherent in graphs. In contrast, Riemannian geometry offers an elegant mathematical framework for modeling structures, while remaining compatible with graph semantic learning, even with LLMs. In this paper, we argue that, for graphs, Riemannian geometry speaks louder than words, and lay out the foundational principles for GFM. Reimagining with Riemannian geometry, we introduce a blue sky idea-Riemannian Foundation Model (RFM)-that opens a new pathway for capturing complex structural patterns and uncovering cross-domain generalities. RFM emphasizes intrinsic graph geometry and embodies endogenous capacities for structural inference and generation, moving beyond mere representation-space switching. Accordingly, we outline a progressive agenda that begins with universal structural understanding through intrinsic geometry, and then rebuilds LLM with a Riemannian engine for general-purpose graph modeling and beyond. Thus, RFM enables a paradigm shift from designing graph models to solving graph-structured applications with RFM agents, unlocking the next-generation graph intelligence.
0
cs.LGcs.AIcs.CL James Wedgwood, Aashiq Muhamed, Mona T. Diab et al. · Mar 23, 2026

Preference alignment typically requires expensive weight-updating training like RLHF or DPO, which lacks mechanistic interpretability. This paper proposes DSPA, an inference-time method that dynamically steers sparse autoencoder (SAE) features based on prompt content without modifying base-model weights. By computing a sparse conditional-difference map $\mathbf{A}$ from preference triples that links prompt features to generation-control features, DSPA edits only token-active latents during decoding. The method achieves competitive open-ended generation quality with up to $4.47\times$ fewer alignment-stage FLOPs than training-based alternatives, while offering direct auditability of which features are modified and revealing that preference directions are dominated by discourse and stylistic signals.

Preference alignment is usually achieved by weight-updating training on preference data, which adds substantial alignment-stage compute and provides limited mechanistic visibility. We propose Dynamic SAE Steering for Preference Alignment (DSPA), an inference-time method that makes sparse autoencoder (SAE) steering prompt-conditional. From preference triples, DSPA computes a conditional-difference map linking prompt features to generation-control features; during decoding, it modifies only token-active latents, without base-model weight updates. Across Gemma-2-2B/9B and Qwen3-8B, DSPA improves MT-Bench and is competitive on AlpacaEval while preserving multiple-choice accuracy. Under restricted preference data, DSPA remains robust and can rival the two-stage RAHF-SCIT pipeline while requiring up to $4.47\times$ fewer alignment-stage FLOPs. Finally, we audit the SAE features DSPA modifies, finding that preference directions are dominated by discourse and stylistic signals, and provide theory clarifying the conditional-difference map estimate and when top-$k$ ablation is principled.
0
cs.LGcs.AIcs.HC Chen Gong, Zhenzhe Zheng, Yiliu Chen et al. · Mar 23, 2026

Machine learning models on mobile devices spend 61-86% of execution time extracting features from user behavior logs rather than running inference. This paper introduces AutoFeature, a graph-based engine that eliminates redundant operations across features and consecutive executions using directed acyclic graph optimization and intelligent caching. Tested across five industrial services including TikTok and e-commerce platforms, it achieves 1.33×-4.53× end-to-end latency reduction without accuracy loss.

Machine learning models are widely integrated into modern mobile apps to analyze user behaviors and deliver personalized services. Ensuring low-latency on-device model execution is critical for maintaining high-quality user experiences. While prior research has primarily focused on accelerating model inference with given input features, we identify an overlooked bottleneck in real-world on-device model execution pipelines: extracting input features from raw application logs. In this work, we explore a new direction of feature extraction optimization by analyzing and eliminating redundant extraction operations across different model features and consecutive model inferences. We then introduce AutoFeature, an automated feature extraction engine designed to accelerate on-device feature extraction process without compromising model inference accuracy. AutoFeature comprises three core designs: (1) graph abstraction to formulate the extraction workflows of different input features as one directed acyclic graph, (2) graph optimization to identify and fuse redundant operation nodes across different features within the graph; (3) efficient caching to minimize operations on overlapping raw data between consecutive model inferences. We implement a system prototype of AutoFeature and integrate it into five industrial mobile services spanning search, video and e-commerce domains. Online evaluations show that AutoFeature reduces end-to-end on-device model execution latency by 1.33x-3.93x during daytime and 1.43x-4.53x at night.