AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

CANN Bench: Benchmarking Agent Generated Kernels against Real NPU and Algorithmic Limits

Xue-Jian Gao, Deng Pan, Yueming Su, Jiasheng Li, Bin Du, Fengming ZHU, Chengdi Ma, Junyi FAN, Qichen Liao, Chengqiu Hu, Xinxian Chen, Lingchao Zheng, Jun Li, Jiwei Yang, Yuwei Fan

AI agents are now capable of writing, compiling, and iteratively optimizing low-level operator kernels on different hardware platforms. Existing benchmarks, however, focus almost exclusively on CUDA and Triton, leaving hardware ecosystems with less-exposed programming models without a common evaluation baseline. We present CANN Bench, an open benchmark for AI-generated operator code on Huawei's Ascend NPU. The current release covers 53 operators and 1060 test cases organized into four difficulty tiers -- from simple elementwise primitives to MoE dispatch and FlashAttention kernels -- spanning FP16, BF16, FP32, and INT8 precision formats. Evaluation adopts a \textbf{three-dimensional weighted composite score} that treats compilation, functional correctness, and performance as independent axes, providing a principled reward signal for kernel-generation agents. Performance is graded against an out-of-the-box PyTorch-on-Ascend baseline and an analytical per-case Hardware-Anchored Performance (HAP) limit on real NPU hardware, ensuring scores reflect genuine optimization headroom rather than measurement artifacts. The evaluation harness is designed to resist reward hacking from the ground up. CANN Bench is versioned within the official CANN repository and is designed for long-term community co-construction, providing the Ascend ecosystem with a quantitative, reproducible, and sustainably maintained yardstick for AI operator-authoring capability.

arXiv

Reward-Aware Population Scaling of Evolutionary Strategies in LLM Fine-Tuning

Sung Cho, Gyubin Han

Using Evolutionary Strategies (ES) for fine-tuning large language models is attractive because it is memory-efficient, parallel, and compatible with black-box or discrete rewards. Yet its population-size conclusions conflict sharply: fine-tuning with cross-entropy (CE) reward succeeds with $N=1$, while binary-reward training often needs $N \approx 30$. We show this gap is largely about reward design and normalization, not population size. In the capable-model regime we study, z-score advantage normalization can cause $N=2$ to fail. Disabling normalization lets binary-reward ES with $N=2$ improve on GSM8K and TREC across capable models spanning 0.5B-7B, where the normalized variant collapses or degrades. This small-$N$ risk is set by reward granularity: binary accuracy reward induces a zero-advantage probability $q$ that depends in closed form on base accuracy, batch size, and intra-pair correctness correlation; a zero-training probe on Qwen2.5-Instruct/GSM8K matches the formula with mean absolute error 0.020 across 12 configurations and finds the availability threshold $N_{\mathrm{avail}}$ to be small in this capable-model regime. The implication is not that $N=2$ is universally sufficient, but that small-population failure in capable-model binary ES can be an implementation artifact rather than an intrinsic population limit.

arXiv

NavEYE: Vision-Centered Multi-Sensor Fusion-Based Situational Awareness System for Intelligent Surface Vehicles

Ryan Wen Liu, Junxiong Liang, Haoyu Wang, Mengwei Bao

With the rapid development of sensor and artificial intelligence (AI) technologies, intelligent surface vehicles (ISVs) have gained increasing attention from academia and industry. Their intelligence, reliability, and safety depend heavily on situational awareness in complex navigational environments. To achieve high-quality perception, we develop a vision-centered multi-sensor fusion system, named NavEYE, by exploiting complementary sensors, including the automatic identification system (AIS), radar, and RGB camera. Specifically, we first propose a multi-constrained gated data association method (MCGA) to accurately match low-temporal-resolution AIS data with high-temporal-resolution radar data. Their fusion result is then obtained by selectively implementing distance-aware adaptively weighted fusion (DAWF) and timeliness decay-based stitching fusion (TDSF), which reduce the uncertainty caused by AIS or radar data loss in real-world sensing scenarios. Based on accurate and robust visual object detection, we further associate and fuse AIS, radar, and visual data through joint constraints of normalized bearing and distance features. According to the fusion results, comprehensive information related to ships of interest can be automatically obtained, helping enhance situational awareness and reduce collision risk for ISVs. The feasibility, robustness, usability, and effectiveness of the proposed multi-sensor fusion method and situational awareness system are demonstrated through extensive experiments on a real-world sensing dataset collected from AIS, radar, and camera. The experimental results show the superior performance of our fusion method in both quantitative and qualitative evaluations. In addition, the shipboard NavEYE system can promote navigational safety for ISVs in complex and dynamic environments.

arXiv

Intelligence-Guided Adaptive Purification for DDoS-Resilient Quantum Networks: A CUDA-Q based Study

Santanu Ganguly

Quantum-repeater networks require adaptive control policies that balance entanglement generation rate, end-to-end fidelity, purification overhead, and memory-induced latency. This tradeoff becomes more complex when the classical control plane is degraded by cyber anomalies or denial-of-service traffic. We develop a CUDA-Q/SeQUeNCe co-simulation workflow for studying adaptive entanglement purification in heterogeneous linear repeater chains. CUDA-Q noisy quantum kernels are used to estimate primitive entanglement purification and swapping behavior, while SeQUeNCe provides an event-layer model for stochastic link generation, waiting-time-dependent memory decay, purification failure, and end-to-end swapping. Under stationary conditions, we compare no purification, local threshold purification, mean-field predictive purification, fixed purification, and a resource-penalized risk-aware predictive policy. In an 8-node chain, the resource-penalized risk-aware controller increases above-target delivery probability relative to fixed purification while reducing latency and purification overhead. We then couple the quantum-network controller to anomaly scores derived from the CSE-CIC-IDS2018 benign-to-SSDP intrusion-detection trace. During the attack period, the attack-unaware controller maintains high raw delivery, but its above-target entanglement delivery falls to 0.098+/-0.007; the IDS-aware resource-adaptive controller switches to more purification-heavy masks and increases above-target delivery to 0.344+/-0.011, closely matching the oracle-aware value of approximately 0.335. These results demonstrate that cyber-state awareness can improve useful quantum-network outcomes by trading raw throughput for fidelity-qualified entanglement delivery.

arXiv

ForensicNet: Lightweight Attention-Enhanced MobileNetV2 for Automated Face Identification

Savitha N J, Lata B T

In forensic environments, automated identification of perpetrators is difficult due to pose changes, changes in light, occlusion, and lack of labeled data. This paper presents ForensicNet, a lightweight deep learning framework for forensic face recognition that enhances attention. The suggested model combines the MobileNetV2 backbone with Convolutional Block Attention Modules (CBAM) to improve the learning of discriminative features while maintaining computational speed. A two-phase transfer learning strategy with adaptive layer unfreezing is used to improve domain adaptation and reduce overfitting. This study used publicly available datasets such as LFW and SCFace, with 15,000 facial images spanning 68 identity classes. The proposed model outperforms baseline architectures such as AlexNet, ResNet-50, and MobileNetV2, with an accuracy of 92.4%, a precision of 90.8%, and a recall of 89.5%. Additionally, the framework requires only 2.1 GFLOPs per inference, and hence can be used in real-time forensic surveillance applications.

arXiv

MedTVL: Harnessing Vision and Language for Medical Time Series Classification

Jiexia Ye, Jia Li, Fugee Tsung

Recent advancements in multimodal learning for medical time series (MedTS) classification highlight the benefits of integrating complementary modalities for clinical decision. However, existing methods typically focus on bi-modal interactions (e.g., time series and text), leaving the tri-modal synergy between time series, vision, and language largely unexplored. Inspired by diagnostic practice synergizing numerical assessment, visual inspection and clinical context, we introduce MedTVL, a text-guided dual-pathway architecture tailored for MedTS classification. Specifically, it synergizes a convolution-based temporal pathway for fine-grained temporal dynamics from raw numerical sequences and a transformer-based visual pathway for holistic morphological structures from time-series-derived images. Such combination of cross-modal and architectural heterogeneity provides a comprehensive diagnostic perspective. To further resolve potential diagnostic ambiguity, both pathways are guided by adaptive medical textual semantics. Finally, a Mixture-of-Experts mechanism dynamically routes each instance to specialized fusion experts, capturing instance-specific reliance on the temporal and visual pathway outputs. In addition, MedTVL supports multimodal contrastive learning to mitigate the clinical label scarcity challenge. Extensive experiments across multiple medical datasets and tasks, spanning supervised, few-shot, and contrastive learning settings, demonstrate the superiority and transferability of MedTVL, highlighting its potential for robust clinical decision support.

arXiv

From Prediction to Intervention: Personalized Meal-Level Glucose Regulation via an LLM Agent

Mingyu Huang, Weiqing Min, Ying Jin, Yilin Wang, Shuqiang Jiang

Personalized glucose regulation remains a central yet unresolved challenge in precision nutrition, as postprandial glucose response varies substantially across individuals. Existing approaches based on glycemic indices fail to adequately account for such heterogeneity and lack the mechanism to dynamically adjust meals based on personal physiological feedback. In this context, recent advances in LLM-based agents offer a promising direction, as they enable context-aware reasoning and iterative refinement. Inspired by this, we propose a physio-feedback agentic loop, a unified system that integrates individualized absorption modeling with dietary intervention to regulate glucose response. Specifically, we develop a Physiology-Aware Glucose Predictor to model individualized absorption dynamics through a learnable Temporal Physiological Absorption Decay Module. We then construct a Prediction-Driven Two-Stage Meal Optimization Agent that iteratively refines real-world meals using predicted outcomes as explicit feedback. Through extensive experiments on multiple public datasets, we demonstrate that our method not only improves prediction accuracy but also effectively reduces glucose excursions. To the best of our knowledge, this paper marks the first step in integrating physiological learning with an LLM-based agent for personalized glucose regulation.

arXiv

RS-Agent: Automating Remote Sensing Tasks through Intelligent Agent

Wenjia Xu, Zijian Yu, Boyang Mu, Jiuniu Wang, Zhiwei Wei, Mugen Peng

Recent advances in Multimodal Large Language Models (MLLMs) have shown promise for remote sensing tasks such as visual question answering and scene understanding. However, existing models remain limited to basic instruction-following and struggle with real-world scenarios that require multi-source data integration, fine-grained spatial reasoning, and domain expertise. To address this gap, we propose RS-Agent, a domain-adapted intelligent agent that connects user intent with professional remote sensing workflows through structured task planning and tool orchestration. RS-Agent consists of four components aligned with typical remote sensing workflows: a Central Controller for intent understanding and process planning, a dynamic toolkit for tool execution, a Solution Space for task-specific expert guidance, and a Knowledge Space for domain knowledge support. We further introduce Task-Aware Retrieval, which improves planning by identifying task types and retrieving expert-defined solutions, and DualRAG, a weighted dual-path retrieval-augmented generation method that enhances the relevance and completeness of retrieved knowledge. RS-Agent natively supports multiple imaging modalities, including optical and SAR imagery, and can automatically organize dedicated SAR processing tools into executable workflows. Experiments on 9 datasets and 18 remote sensing tasks show that RS-Agent significantly outperforms state-of-the-art MLLMs, achieving over 95% task planning accuracy and strong results in scene classification, object counting, and remote sensing visual question answering. These results demonstrate the value of combining LLM reasoning with remote sensing expertise for intelligent geospatial analysis.

arXiv

InfoDPP-PAC: Principled Patch Selection for Whole Slide Image Analysis

Prateek Mittal, Ayush Srivastava, Joohi Chauhan

Each WSI slide contains thousands of candidate tissue patches, while supervision is usually available only at slide level. Existing bag-construction strategies like Uniform extraction and handcrafted heuristics do not control redundancy while attention-based multiple-instance models couple patch importance to a particular downstream classifier, and coreset methods optimise embedding-space coverage without modelling task-relevant patch quality. We introduce InfoDPP-PAC, a principled patch-selection framework that combines teacher-seeded Gaussian process relevance modelling, determinantal log-determinant diversity,submodular greedy optimisation, and a concentration-based adaptive stopping rule. The main theoretical result shows that the log-determinant diversity term used in DPP-style selection is the Gaussian process mutual information between a selected subset and the latent relevance function. We further derive a PAC-style certificate for residual information gain, allowing the number of retained patches to vary by slide rather than being fixed a priori. The empirical study evaluates whether the selected subset is diverse, spatially and morphologically covering, non-redundant, and enriched for the teacher-derived relevance signal. It does not claim end-to-end diagnostic improvement after retraining a downstream MIL model. On 202 HISTAI gastrointestinal whole-slide images, the adaptive rule uses 83.7% fewer patches on average than a fixed full budget while retaining 97.9% of full-budget composite selection quality. At a matched budget, InfoDPP-PAC achieves the highest mean teacher-derived relevance score among fourteen baselines, with diversity and composite scores close to the strongest coreset methods. The results support InfoDPP-PAC as a controlled quality-diversity-cardinality selection framework, rather than as a downstream clinical predictor.

arXiv

An adaptive and evolvable deep reinforcement learning framework for weather prediction

Qiang Wu, Han Li, Jianping Huang

No single AI weather model excels at all variables, pressure levels, and lead times. Rather than building yet another architecture, we reframe the forecasting problem as one of coordination. Here we present Feitian Adaptive Ensemble Weather (FTAE-Weather), a lightweight framework that learns, through deep reinforcement learning, when and where to trust each member of an open pool of pretrained forecasters. A tactical Weight-Agent reads the current atmospheric state and assigns variable- and horizon-specific fusion weights, while a strategic Evolve-Agent periodically prunes underperforming models and absorbs newly released ones. Asynchronous prediction caching keeps training cost independent of the slowest constituent model. Adding fewer than 0.01 percent extra parameters, FTAE-Weather reduces RMSE by from 17.2 percent to 78.3 percent over the best individual model in 10 atmospheric variables and outperforms conventional ensemble baselines across lead times from 72 to 360 hours. The framework thus converts a growing, fragmented inventory of specialist models into a single prediction system that strengthens as the field of AI weather forecasting releases new architectures-turning model diversity from a coordination challenge into a compounding scientific advantage.

arXiv

Self-Generated Text Recognition: Quality Heuristics, Cross-Task Transfer, and Downstream Bias in LLM Evaluation

Jesse St. Amand, Callum Canavan, Sohaib Imran, Joseph Hewson, Aaron Lutz, Shi Feng, Puria Radmard, Lennie Wells

Self-Generated Text Recognition (SGTR)--the ability of an LLM to identify its own outputs--poses risks to AI safeguards that rely on LLMs as evaluators or monitors: an LLM may recognize outputs from other copies of the same model and make biased judgments or collude outright. Prior work has drawn conflicting conclusions about whether current models possess significant SGTR capabilities. We explain these disagreements by identifying key experimental design choices--which we term operationalizations--that drive divergent results. Evaluating 13-21 models across six presentation operationalizations and four task-domain operationalizations, we find that accuracy varies substantially with evaluation format (pairwise vs individual assessments of text), conversation format (presenting candidate text in user tags vs assistant tags), and the domain of the task used to generate candidate text (e.g., coding vs summarization). We corroborate previous observations that a quality heuristic--models attributing authorship to text they perceive as higher quality--is a dominant confound. We also find that improving a model's SGTR performance via supervised fine-tuning (SFT) on one operationalization can generalize to others, and can increase the model's preference for its own outputs when it acts as a judge in the AlpacaEval framework. Our results suggest that, despite confounds, some models possess practical SGTR capabilities, and that SGTR should be monitored and considered in the design of safety-critical AI applications.

arXiv

MAME: Multidimensional Adaptive Metamer Exploration with Human Perceptual Feedback

Mina Kamao, Hayato Ono, Ayumu Yamashita, Kaoru Amano, Masataka Sawayama

Alignment between human brain networks and artificial models has become an active research area in vision science and machine learning. A widely adopted approach is identifying "metamers," stimuli physically different yet perceptually equivalent within a system. However, conventional methods lack a direct approach to searching for the human metameric space. Instead, researchers first develop biologically inspired models and then infer about human metamers indirectly by testing whether model metamers also appear as metamers to humans. Here, we propose the Multidimensional Adaptive Metamer Exploration (MAME) framework, enabling direct, high-dimensional exploration of human metameric spaces through online image generation guided by human perceptual feedback. MAME modulates reference images across multiple dimensions based on hierarchical neural network responses, adaptively updating generation parameters according to participants' perceptual discriminability. Using MAME, we successfully measured multidimensional human metameric spaces within a single psychophysical experiment. Experimental results using a biologically plausible CNN model showed that human discrimination sensitivity was lower for metameric images based on Gram-matrix representations derived from low-level CNN features than for those derived from high-level CNN features. The finding suggests a relatively worse alignment between the metameric spaces of humans and the CNN model for low-level processing compared to high-level processing. Counterintuitively, given recent discussions on alignment at higher representational levels, our results highlight the importance of early visual computations in shaping biologically plausible models. Our MAME framework can serve as a future scientific tool for directly investigating the functional organization of human vision.

arXiv

Jais 2: A Family of Arabic-Centric Open Large Language Models

Mohamed Anwar, Abed Alhakim Freihat, George Ibrahim, Mostafa Awad, Abdelrahman Sadallah, Gurpreet Gosal, Gokulakrishnan Ramakrishnan, Sarath Chandran, Biswajit Mishra, Rituraj Joshi, Ahmed Frikha, Etienne Goffinet, Abhishek Maiti, Ali El Filali, Sarah AlBarri, Samujjwal Ghosh, Rahul Pal, Parvez Mullah, Awantika Shukla, Sajid siddiki, Samta Kamboj, Onkar Pandit, Sunil Kumar Sahu, AbdelRahman Elbadawy, Amr Mohamed, Ahmad Chamma, Evan Dufraisse, Abdelaziz Bounhar, Dani Bouch, Hadi Abdine, Guokan Shang, Fajri Koto, Yuxia Wang, Zhuohan Xie, Ali Mekky, Rania Elbadry, Sarfraz Ahmad, Momina Ahsan, Omar El Herraoui, Daniil Orel, Hasan Iqbal, Kareem Elzeky, Mervat Abassy, Kareem Elozeiri, Saadeldine Eletter, Farah Atif, Nurdaulet Mukhituly, Haonan Li, Xudong Han, Aaryamonvikram Singh, Zainul Abedien Ahmed Quraishi, Neha Sengupta, Larry Murray, Avraham Sheinin, Joel Hestness, Natalia Vassilieva, Hector Xuguang Ren, Zhengzhong Liu, Michalis Vazirgiannis, Preslav Nakov

Jais 2 is a family of Arabic-centric large language models developed jointly by MBZUAI, Cerebras, and Inception, designed to advance Arabic-centric language modeling, with strong performance across the Arabic and culturally grounded benchmarks evaluated in this report. The family includes, to our knowledge, the largest open Arabic-centric LLM trained from scratch at 70B parameters, and a competitive 8B-parameter variant among the evaluated open models. A custom Arabic-centric vocabulary enables efficient training and inference. In addition, an optimized architecture and training recipe yield highly compute-efficient training. With a substantially smaller token budget than comparable models, Jais 2 achieves strong Arabic performance on the benchmarks considered in this report and competitive English results. The models obtain leading results among the evaluated open models on OALL2 and AraGen. They also perform strongly on several culturally grounded Arabic benchmarks, including poetry, religion, cuisine, and dream interpretation, as well as in general tasks such as translation and summarization. We release the models in HuggingFace under a commercially permissive license. Jais 2 70B is also released as a chat app on the Web, iOS, and Android; it runs on Cerebras hardware, delivering up to 2,000 tokens per second, and enabling high-throughput Arabic-centric chat serving in our deployment setting. By uniting scale, linguistic diversity, cultural fidelity, openness, and speed, Jais 2 provides an open-weight foundation intended to support further research and development in Arabic-centric LLMs.

arXiv

Geographically Weighted Surrogate Models for Rapid Small-Area Chronic Disease Estimation

Aanya Gupta, Szandra Péter, Sara Von Hoene, Emma Von Hoene, Taylor Anderson

Small-area estimation (SAE) enables researchers and policymakers to identify spatial disparities in health outcomes, but survey-based SAE products carry an inherent lag. Gold-standard estimates such as CDC PLACES are released roughly two years after the underlying survey data are collected, limiting their use for time-sensitive decision-making. This study evaluates the potential for machine learning (ML) to serve as a surrogate, learning the relationship between frequently updated area-level predictors and existing SAE outputs to generate timely, comparable estimates in years when SAE from surveys are unavailable or delayed. We evaluate several global and geographically weighted ML models for county-level SAE of ten chronic conditions across the US: COPD, asthma, heart disease, arthritis, cancer, depression, diabetes, high blood pressure, high cholesterol, and stroke. Our findings suggest that geographically weighted ML frameworks like geographically weighted random forest and geographically weighted regression offer scalable and open data surrogates for rapidly generating SAE and supporting data driven decision making.

arXiv

HoloCount: A Holistic Visual Counting Benchmark for MLLMs

Jinhong Deng, Limeng Qiao, Guanglu Wan

Visual counting is a fundamental pillar of multimodal intelligence, requiring a seamless integration of fine-grained grounding and spatial reasoning. While Multimodal Large Language Models (MLLMs) have achieved remarkable success in qualitative scene understanding, their quantitative precision remains a significant bottleneck, often characterized by persistent numerical hallucinations. Existing counting benchmarks primarily focus on basic perception in simplified contexts, failing to capture the complex failure modes that emerge under logical constraints or adversarial conditions. To address these limitations, we introduce HoloCount, a holistic and diagnostically rich benchmark structured around a three-level hierarchical taxonomy. HoloCount evaluates MLLMs across: (1) Semantic Counting, focusing on atomic and property-based enumeration; (2) Analytical Counting, assessing logical composition through spatial and set-based reasoning; and (3) Robustness Testing, probing model integrity against adverse scenarios and grounded counter-priors, such as high-density scenes and linguistic biases. Through an exhaustive evaluation of over 20 state-of-the-art MLLMs, we reveal a critical performance gap: even top-tier models degrade significantly as tasks transition from perception to complex analytical reasoning and adverse scenarios. Our findings provide a systematic landscape of current MLLM counting capabilities and offer a roadmap for developing more grounded and reliable multimodal systems. The dataset is available at https://mm-mvr.github.io/HoloCount/.

arXiv

Psych-ECA: A Reproducible Semi-Synthetic Benchmark for Synthetic Control Arms in Longitudinal Psychiatry

Aakash Bhagat, Shashank Choudhary

External and synthetic control arms (ECAs) are entering psychiatric drug development, but the field lacks a benchmark that evaluates the properties regulators care about: not only how accurately a method reconstructs untreated trajectories, but whether its uncertainty is calibrated, whether it is robust to the informative observation times common in mental-health records (sicker patients are seen more often), and what false-positive rate it induces in go/no-go trial decisions. Real psychiatric trial data (e.g. STAR-D and registry cohorts) require credentialed access and lack ground-truth counterfactuals, so, following established semi-synthetic benchmarks in causal inference (IHDP, ACIC, and the PK-PD tumor-growth simulator), we release Psych-ECA, a fully reproducible generator of longitudinal symptom trajectories for depression (PHQ-9), anxiety (HAM-A), and psychosis (PANSS) with known counterfactual control arms, informative visits, and validated-scale measurement noise. We benchmark eight estimators spanning carry-forward, pooled real-world-data averages, nearest-neighbour matching, linear mixed models, gradient boosting, and the Scribe trajectory-bridge method. Three findings emerge. First, trajectory and flexible machine learning methods achieve the best counterfactual accuracy (about 2.3 PHQ-9 RMSE), outperforming cross-sectional baselines. Second, only Scribe is both accurate and calibrated, achieving 93-96% empirical coverage of nominal 90% prediction intervals, compared with 87-88% for gradient boosting and 62-75% for uncalibrated SDE models. Third, inverse-intensity correction reduces bias under informative sampling, while Scribe's calibrated intervals are the only trajectory method that maintains nominal false-positive rates as informativeness increases. We release all code, data-generation scripts, and random seeds to enable fully reproducible evaluation.

arXiv

Discrete Action Space as a Prerequisite for GRPO Convergence in Small-Model Continuous Control

Dmytro Filatov, Valentyn Fedorov, Vira Filatova

We study whether Group Relative Policy Optimization (GRPO) can fine-tune small language models for simulated quadrotor continuous-control tasks. In our benchmark, vanilla GRPO fine-tuning of Qwen-0.5B for 25 Hz quadrotor velocity control collapses to the trivial zero action: 0 percent success rate, with entropy falling from 0.35 to 0.03 within 60 steps. Two ablations - removing the jerk-penalty term and removing the KL anchor to the pretrained prior - each prevent entropy collapse, yet neither enables learning. When the action interface is replaced by a 5-way categorical choice over PID presets, training converges. The resulting controller traces a smoothness-reliability Pareto frontier along training duration; both endpoints are reported: 98.6 percent success with 0.656 m/s3 jerk at 64 steps, and 100 percent success with 1.103 m/s3 jerk, or 0.796 under a matched velocity cap, at 256 steps. The recipe is evaluated across three pretrained language models. As context, a re-tuned classical baseline, PID with Ki = 0.30 and vmax = 2.5, reaches the same 100 percent success rate at jerk 0.736 m/s3. A high-fidelity simulation using Crazyflie 2.1 dynamics surfaces a hover-region training-distribution gap.

arXiv

Trajectory-Aware Retrieval Agents for Temporal Decision- Making

Jing Wang, Jie Shen, Xing Niu

We study the problem of decision-making from long-form, temporally structured text using large language model (LLM) agents. Standard retrievalaugmented generation (RAG) pipelines fragment chronological context into isolated snippets, discarding the temporal structure that is often critical for correct downstream decisions. We introduce TLM (Trajectory Language Model), a closed-loop agentic framework that iteratively refines the evidence set using SHAP-guided feedback. The key technical contribution is the latent growth curve model (LGCM) over retrieved chunk embeddings, which provides an interpretable mechanism for detecting trajectory trends, turning points, and information gaps. We show that, under a scorer-calibration assumption (which holds approximately in practice), the iterative refinement procedure is monotonically non-decreasing in the probability assigned to the correct label. Empirically, TLM is evaluated on three temporally grounded decision tasks: medical question answering, earnings call surprise prediction, and overnight stock gap prediction. TLM substantially outperforms both zero-shot LLM baselines and standard retrieval-augmented approaches on the medical task, and yields consistent, economically meaningful gains on the two financial tasks.

arXiv

Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems

Xiaoyang Cao, Siddarth Srinivasan, Michiel A. Bakker

End-to-end reinforcement learning can improve the accuracy of compound LLM systems, but it does not constrain how modules divide labor internally. We identify Role Drift, a failure mode in which modules preserve or improve end-task performance while deviating from their assigned roles through role-violating shortcuts that remain invisible to system-level evaluation. To make role drift observable and controllable, we propose Role Anchor, a regularizer that modulates how much each module deviates from its assigned role during end-to-end training. The key idea is to preserve how the role prompt shifts the module's next-token predictions relative to a neutral prompt, which serves as a proxy for the role's intended effect during training. Experiments on two compound LLM pipelines reveal role drift that accuracy alone fails to detect: a decomposer meant to split a question into sub-questions for a separate solver instead plants the answer in them, and a reader meant to answer from retrieved passages instead falls back on parametric memory. In fact, on the decomposer pipeline this shortcut drives most of the apparent RL gain: 86% of it vanishes once the decomposer is held to its role, indicating that terminal accuracy alone can badly overstate how much a compound system has genuinely learned. Across both pipelines, Role Anchor mitigates role drift at a tunable accuracy cost that varies by pipeline and anchor strength. Additional gradient analysis suggests that the regularizer reduces alignment with the role-drift direction rather than simply suppressing learning.

arXiv

FBLayout: Optimizing Memory Layout for Efficient LLM Finetuning on Mobile GPUs

Kahou Tam, Wei Niu, Yu Bao, Xiaomin Ouyang, Chengzhong Xu, Li Li

Transformer-based models have enabled unprecedented capabilities across language, vision, and multimodal tasks. On-device fine-tuning of transformer models offers a privacy-preserving path to personalized AI, yet remains inefficient on mobile GPUs due to severe memory constraints and frequent layout transformations in attention mechanism during training. Existing mobile training frameworks either use unified layouts for forward and backward passes -- leading to fragmented memory access and poor GPU utilization during backpropagation -- or rely on explicit layout conversions, which introduce significant transformation overhead. To overcome this, we propose FBLayout, a layout-aware framework that co-designs tensor organization with mobile GPU platforms. FBLayout introduces: (1) a unified R-Tile layout for multi-dimensional reductions across forward/backward passes; (2) tile-based index transformation to eliminate physical data movement; and (3) activation-guided layout selection to propagate efficient layouts globally. Evaluations on seven transformer models across different mobile phones (including ARM Mali and Qualcomm Adreno GPUs) show that FBLayout achieves 2.2-5.7x speedup over MNN, TFLite, and TVM, while significantly improving cache efficiency and reducing memory footprint, enabling practical on-device large model fine-tuning.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.