AI agents and reinforcement-learning (RL) trainers run thousands of short, isolated GPU jobs: tool calls, verifiers, and rollouts. A job often finishes in seconds, but starting a GPU pod for it takes 20-35 s even when the container image is already on the node, and minutes when it is not. A warm pod removes the wait only if it can serve the next job, and keeping pods warm must neither waste GPUs nor leak state between callers. Firepool, the sandbox and execution broker behind agentic and RL workloads at LinkedIn, meets both requirements. It makes pods interchangeable by building them from a shared base image per hardware profile and binding each job's code, extra dependencies, and cluster topology when the job starts. It keeps them warm cheaply and safely by reusing a pod only for the caller that used it last, staging zero-VRAM standby pods in fractional GPU shares alongside active pods, and capping the GPUs that warm pools may hold via queue-driven budgets. Over a sampled peak period in production, Firepool serves 94.7% of 3,206 pod requests from warm or standby capacity. In paired benchmarks with fixed-duration jobs on production H200 nodes, agent and RL loops finish 13x and 10.8x faster than with a fresh pod per job; against a fresh reuse tag per job, warm reuse cuts p90 start latency from 59.1 s to 3.5 s. When a job cannot reuse the previous pod, standby promotion shortens the handover wait from 51 s to 0.94 s. In local and simulated runs, a keep-alive matched to the loop's pauses keeps the latency benefit while holding about half the capacity of an always-on pool.
Vision-Language Models have had massive leaps in recent years. Yet, when tested on high-level cognitive tasks like pattern completion, abstract logic, and relational reasoning, they still fall short. This limitation stems from the models' lack of ability to find out underlying rules and maintain constant intermediate steps, rather than merely deciding on the correct option. To research these shortcomings, this study analyzes visual reasoning performance throughout a multilingual evaluation framework encompassing the B\.ILSEM, MM-IQ, Visual Puzzles, and PuzzleVQA datasets. We evaluate various data mixing strategies and supervised fine-tuning (SFT) methods. We observed that a correct answer does not always mean the reasoning chain is valid. Because of this, we adopt an LLM-as-a-Judge approach to assess the quality of reasoning beyond final answer accuracy. Our experimental results reveal that baseline models often exhibit more stable reasoning compared to fine-tuned counterparts. Most importantly, we observe that SFT can lead to model collapse, manifesting as infinite text loops or empty outputs. The study concludes that current VLMs still struggle with abstract visual reasoning.
High-level synthesis (HLS) can translate C/C++ kernels into register-transfer level (RTL) designs, yet generating high-quality hardware still relies heavily on manual pragma tuning and repeated synthesis evaluation. We propose AutoC2V, a C++-to-Verilog generation framework driven by real synthesis feedback that enables large language models (LLMs) to directly generate RTL with improved quality of results (QoR) from synthesizable C++ kernels. We construct training data from real synthesis results, including synthesizable C++ source code, the corresponding Verilog RTL, QoR metrics, and optimization objectives. Using this dataset, we align the model toward synthesis-aware RTL generation through Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). To reduce the high cost of repeatedly invoking real synthesis during GRPO, we design the QoR Prediction Reward Model (QoR RM), which predicts latency and resource metrics by encoding RTL into long-context semantic representations, thereby providing low-cost and continuous QoR reward signals for policy optimization. Experimental results demonstrate that AutoC2V not only substantially improves RTL functional pass rates, but also achieves superior latency and resource optimization compared to the unfine-tuned base model, representative HLS optimization approaches (HLSPilot, LLM-DSE, and AutoDSE), and advanced general-purpose commercial models (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, DeepSeek v4 Pro, and GLM-5.2). These results indicate that combining real synthesis data with efficient QoR reward modeling can further improve the ability of LLMs to generate high-quality RTL.
Iterative neural policies usually spend the same computation on every decision, although the benefit of additional processing varies across states. We introduce Adaptive Internal Computation (AIC), an iterative policy that learns when to stop refining its latent representation and act. In AIC, we train action selection and stopping jointly, an expected-depth penalty values computation and intermediate value predictions provide auxiliary supervision. We show that when only some states benefit from further refinement, an ideal state-dependent allocator outperforms every state-independent stopping schedule at the same expected depth. With at most eight refinement steps per decision, AIC executes 3.3 steps on average across twelve benchmark tasks and lowers per-decision latency by 31--44% relative to a fixed eight-step policy on four control and robotics tasks. Its mean reward exceeds the best fixed-depth policy on seven tasks and trails it on four. Ablating adaptive stopping and value supervision separately shows that stopping accounts for most of the reward gain on PickCube and AntMaze, where it also executes more refinement steps and that supervision accounts for most of it on RareEvent.
Recent progress in vision foundation models (VFMs) has made automated open-vocabulary 3D bounding box annotation increasingly feasible by jointly leveraging LiDAR point clouds and images. However, partial observations remain a major challenge in sparse driving scenes, where visible LiDAR points often cover only a small fraction of the object's surface and provide insufficient cues for estimating its full spatial extent. As a result, existing methods often produce mislocalized boxes or leave objects unlabeled. To address this limitation, we propose \textbf{\Ours{}}, an automated annotation framework that leverages image-generated 3D assets as amodal geometric priors for partially observed objects. By establishing correspondences between input images and rendered views of 3D assets, \Ours{} robustly aligns canonical amodal geometry with partially observed world-space LiDAR point clouds. To improve temporal consistency and reduce computational cost, \Ours{} generates a single asset per instance, estimates its metric scale once at a single frame and propagates it across frames for dynamic instances, and aggregates multi-frame observations for static instances. 3D object detectors trained on annotations from \Ours{} improve performance over strong baselines by 14.1 mAP on the nuScenes dataset and by 12.45 AP at IoU 0.5 for the vehicle class on the Waymo Open Dataset (WOD). Our project page is available at: \href{https://oliver0922.github.io/AmodalBox/}{https://oliver0922.github.io/AmodalBox/}
Real-time cooperative perception (CP) in multi-agent traffic environments is highly resource-intensive, which posing scalability challenges for large-scale deployment. Efficiency can be significantly improved by selectively performing CP only in safety-critical scenarios and for key occluding objects. Vision-Language Models (VLMs) are well-suited for complex reasoning in dynamic and uncertain environments, making them a promising tool for this selective CP task. This study proposes a VLM-driven agentic framework for blind spot reasoning and dialogue-based CP. A mixed-model architecture is designed that integrates conventional object detection for stable perception, vision-language reasoning for identifying impactful occlusions, and a physics-based geometric module to align reasoning with real-world visibility constraints. The framework also incorporates a Chain-of-Thought (CoT) prompting mechanism to enhance interpretability and intent-aware decision-making. We collect real-world data from connected vehicles operating in urban traffic and compare the driving-critical blind spot identification and dialogue performance of four state-of-the-art VLMs. Experimental results demonstrate that our method achieves over 80% accuracy in driving-critical blind spot detection and over 90% success in dialogue-driven perception sharing. This work provides a scalable, interpretable, and bandwidth-efficient alternative to conventional CP pipelines, and offers new insights into agentic communication design for future connected traffic systems.
In this work we proposed a architecture designed for the classification of ASD in toddlers using eye-tracking data, addressing the critical need for early and precise detection. The proposed model incorporates FMMix, a feature-level augmentation technique that bolsters the model's learning capacity, and in this work we proposed the ATHENA, which captures intricate feature relationships via an adaptive hypergraph framework. Evaluated on two datasets dataset of gaze coordinates-the model achieves state-of-the-art performance, with AUC scores of 99.80% and 99.86%, respectively. It also exhibits high accuracy of 98.45% and 98.56%, sensitivity 98.01% and 98.10%, and specificity 98.74% and 99.02%, respectively. Demonstrating its superior ability to identify subtle gaze patterns associated with ASD, these results significantly surpass existing methods, highlighting the model's potential for clinical ASD screening. Future efforts will focus on validation across broader populations and integration with multi-modal data to further refine diagnostic accuracy.
Video fusion aims to integrate complementary cues from heterogeneous sources, producing visually informative videos with coherent spatio-temporal structures. Jointly learning cross-modal complementarity and temporal dependencies typically relies on large-scale, high-quality video datasets, whose storage and repeated use during optimization incur substantial costs, making data-efficient learning highly desirable. To address this challenge, we rethink video fusion from the perspective of dataset condensation and propose STAR-VF, the first framework that condenses spatio-temporal representations for data-efficient video fusion. STAR-VF condenses the original training set into a compact set of learnable spatio-temporal representations, which are decoded into synthetic video clips and optimized using feature-level supervision provided by a frozen proxy model. The resulting condensed data preserves fusion-relevant spatial and temporal cues without requiring ground-truth fused videos as supervision. To adapt fusion learning to condensed representations, we equip the final fusion model with flow-free inter-frame temporal modeling and query-driven intra-frame aggregation, enabling direct temporal interaction and adaptive selection of fusion-relevant information in representation space. Extensive experiments on infrared-visible, multi-exposure, multi-focus, and medical video fusion benchmarks demonstrate that STAR-VF, trained with only approximately 1/16 of the original temporal training set, achieves fusion quality comparable to methods trained on the full data. Further analyses validate the contributions of learned condensation and flow-free temporal modeling, and demonstrate the robustness of the condensed supervision across different proxy encoders and downstream fusion models. Project page: https://zju-xyc.github.io/STAR-VF-Project-Page/.
Individuals experiencing suicidal ideation (SI) or psychological distress may disclose their social identities when seeking support from LLMs. Attending to identity-related stressors can be important for providing contextually appropriate and empathetic support. However, disclosing an identity does not necessarily mean that it is a cause of distress. In this work, we audit how LLMs interpret distress through disclosed identities. Using paired counterfactual interventions, we evaluate six LLMs across two SI and distress datasets, 16 identity and control conditions, and two prompting protocols. Our findings suggest that models can transform identity disclosures into explanatory narratives, attributing distress to presumed identity-related adversity. These assumptions frequently appear within otherwise empathetic responses, particularly when marginalized identities are disclosed. We contextualize our findings through the lens of compassionate othering to characterize these patterns as a form of harm in which empathetic responses impose identity-linked assumptions on individual suffering. These responses can misrepresent individuals' experiences by treating their identities as explanations for distress without supporting evidence. Our findings highlight the need to mitigate these harms by grounding interpretations of distress in individuals' disclosures and avoiding unsupported identity-linked assumptions.
Hierarchical Joint-Embedding Predictive Architectures (H-JEPA) enable long-horizon visual control but rely on costly test-time gradient descent for planning. Our analysis reveals that iterative optimization consumes up to 97.7% of episode runtime and that low predicted rollout costs do not necessarily translate into successful control. We further identify a planning-horizon asymmetry between long-horizon top-level composition and short two-step subordinate planning. To address these challenges, we propose WARP-JEPA, a hierarchical amortized planning framework that replaces online gradient optimization with learned multi-candidate trajectory proposers while keeping the world models frozen. WARP-JEPA combines a lightweight world-model-in-the-loop causal Transformer at the top level with parallel residual MLPs at subordinate levels. The proposers are trained using model-based rollout costs and best-of-K behavior cloning, enabling efficient candidate selection without test-time gradient optimization. Across 30 frozen H-JEPA models on four benchmarks, WARP-JEPA achieves 9.6-18.8x faster episode inference on long-horizon tasks (2.9-3.0x on Cube) and improves task success by up to 33.3 percentage points. These results demonstrate that amortized planning can substantially improve both inference efficiency and control performance without retraining the world models. Code is available at https://github.com/hideinsewer/WARP-JEPA.
Bilevel optimization addresses machine learning problems in which the outer objective is a value function implicitly defined through the solution of an inner problem. While its algorithmic aspects are well developed, its generalization properties remain much less understood. In this paper, we study the generalization of inexact gradient methods when both objectives are least-squares regression problems and when the estimator of the value function's gradient, or hypergradient, is obtained by implicit differentiation in function space. This approach allows us to derive stationarity bounds for these methods in terms of the uniform generalization errors of the two regression functions needed to estimate the hypergradient. When these regression functions are defined as overparameterized two-layer ReLU networks trained by gradient flow with early stopping, we obtain explicit finite-sample bounds on these generalization errors in the Neural Tangent Kernel (NTK) regime. Combining these with our general inexact gradient analysis shows the effect of sample size on the overall convergence of the bilevel method. Numerical experiments on synthetic data illustrate our results. Beyond the NTK regime, further experiments reveal an initialization-dependent implicit bias induced by the bilevel structure, which can lead to feature learning at the inner level and overall performance improvements.
Hybrid quantumclassical models are increasingly proposed for nancial prediction, but reported gains are rarely separated from the classical components around them. We study a hybrid quantum temporal convolutional network for daily cross-sectional ranking of 2000 Tokyo Stock Exchange equities under a protocol built to make the quantum contribution identiable: six conditions (circuit trained, frozen or absent, a parameter-matched classical surrogate, and the circuit alone, trained or frozen) through one code path; four walk-forward folds with ve seeds; bootstrap intervals, Deated Sharpe Ratios, backtest-overtting probabilities and family-wise error control; ve classical baselines including a gradient-boosted tree; and a locked holdout. On 500 out-of-sample days two one-sided tests place the trained-versus-frozen dierence inside 0.03 annualised Sharpe (p = 0.0055, minimum detectable eect 0.02): the circuit is equivalent to a frozen one, not merely indistinguishable from it. With the circuit as the head's only input the two are equivalent inside 0.25; against the network with no circuit the design resolves only 0.50. No condition is distinguishable from zero (best De-ated Sharpe Ratio 0.17). The holdout reproduces the ablation on a protable window, and 2000 US equities reproduce it where every condition loses money.
We improve Turkish mathematical reasoning in Large Language Models by combining LoRA fine-tuning, weight-space model merging, and search-based inference. Using Turkish-Llama 8B-DPO with Turkish translations of MetaMath, NuminaMath, and TigerLab, we find data quality beats scale (Mix-30k best among fine-tuned; large MetaMath underperforms). Linear merging delivers the largest gains (GSM8K up to 0.63), while Ties is modest and Breadcrumbs/Slerp degrade. Search with MCTS and Tree of Thoughts improves mean LLM-judge scores for reasoning quality on a 0-100 scale. Limitations include machine translation, limited scale, and search stochasticity.
Hardware designs may be available only as gate-level netlists or input-output descriptions when the original register-transfer-level (RTL) design and documentation are not accessible. Low-level representations obscure word-level computation and its finite-width behavior. Recovering high-level RTL from such representations therefore helps engineers understand a circuit's functionality and enables RTL rewriting and resynthesis. A plausible function model is also insufficient unless the generated RTL is equivalent to the original circuit for all input combinations. In this paper, we present RTLREX, a large language model (LLM)-guided framework that coordinates domain-specific tools for function recovery, RTL generation, independent verification, and cost-guided rewriting. Independent verification establishes equivalence between the recovered RTL and the original circuit. The LLM uses tool feedback to select further analyses, revise candidate RTL, and propose syntax-cost reductions. Experiments on 30 benchmarks from Problem A of the ICCAD 2022 CAD Contest show that RTLREX recovers equivalent high-level RTL for all benchmarks, surpassing the state-of-the-art method WolFEx in recovery coverage. On benchmarks recovered by both methods, RTLREX reduces mean syntax cost by 20.5% relative to WolFEx. On complete truth-table benchmarks from the IWLS 2026 Contest, our framework reduces mean syntax cost by 95.7% relative to an LLM agent baseline on the benchmarks validated by both methods.
This study investigates sequential multi-task learning using Parameter-Efficient Fine-Tuning (PEFT), LoRA, and 4-bit quantization on Llama 3.1 8B Instruct for mental health problem detection, emotion detection, and sentiment analysis. Using 4,243 English social media posts, it compares two task sequences and demonstrates that task ordering significantly affects knowledge transfer and task interference. The MHP-first sequence achieves the best mental health problem detection (Micro-F1: 0.7624), while the Emotion-first sequence performs better in emotion (F1: 0.6052) and sentiment analysis (F1: 0.9459). A Joint-MTL baseline further highlights the challenges of simultaneous learning, including task interference and generative format collapse. Overall, the findings demonstrate the effectiveness of sequential PEFT for mental health NLP while showing that task order is critical to balancing performance across related tasks. The integration of LoRA and 4-bit quantization also reduces computational and memory requirements, supporting the adaptation of large language models in resource-constrained environments.
Parameterized instantaneous quantum polynomial-time (IQP) circuits generate distributions conjectured to be hard to sample classically. As reinforcement learning (RL) policies, their $2^n$ measurement outcomes are typically mapped to a smaller action space by a classical decoder. This aggregation can mitigate initialization concentration while making the policy classically simulable. We characterize this tradeoff through the decoder's Fourier--Pauli expansion. Under independent uniform initialization, we express policy variance as an exact sum over groups of nonconstant Pauli components with identical active-generator signatures. Each group contributes its squared aggregate coefficient times $2^{-r}$, where $r$ is the binary rank of its active generators. An inverse-polynomial aggregate coefficient with logarithmic rank prevents exponential policy concentration. Sparse Pauli expansions and contiguous action bins permit efficient classical additive estimation of action probabilities. We introduce a majority decoder combining protected low-weight components with exponentially large exact Fourier support, preserving initialization nonconcentration while excluding polynomial-size enumeration of the exact expansion. Numerical experiments corroborate the variance laws. In ten-qubit CartPole experiments, majority achieves higher returns than recursive parity and the tested local readouts, while higher-weight Pauli sectors contribute an increasing relative share of gradient magnitude during training. A complementary 21-qubit block-chain policy outperforms an extensive pool of fully connected classical neural networks in episode efficiency, even though its bounded-width structure permits efficient classical simulation. These results guide decoder design for quantum and quantum-inspired RL, distinguishing policy nonconcentration, learning performance, and classical simulability.
Zeroth-order (ZO) optimization estimates descent directions from function values alone, which removes the activation storage required by backpropagation and makes it attractive for memory-efficient fine-tuning of large language models (LLMs). Existing low-rank ZO methods either parameterize the additive update through Euclidean factors with a non-unique representation, or restrict each finite-difference direction to be low-rank. In this work, we propose to constrain each additive update block to a fixed-rank matrix manifold while keeping the pretrained weights frozen, so the learned update remains rank-constrained throughout fine-tuning. On this geometry, we build a Riemannian ZO gradient estimator that projects a random matrix onto the tangent space, normalizes it to correct the scale mismatch induced by projection, and further propose Zeroth-order Riemannian Gradient Descent (ZO-RGD) and its smoothed variant, ZO-RGD-S. We show that the ambient finite difference matches its retracted counterpart, the normalized estimator recovers the Riemannian gradient in expectation, and ZO-RGD converges under trajectory-regularity conditions. Experiments on LLM fine-tuning tasks show that the proposed algorithms are competitive with strong forward-only baselines across model families and tasks, with gains in several settings.
Under strict privacy regulations in the clinical domain, de-identification is necessary to enable the use of clinical data such as for natural language processing. However, we currently lack sophisticated ways to tell which de-identification models actually are suitable. Specifically, we need benchmarks that are challenging, diverse, and clinically plausible, all while ensuring label correctness. To this end, we present SPARCLING, an LLM-based pipeline to generate synthetic discharge letters with character-exact PHI span annotations. Our pipeline combines structured attribute sampling, template generation, and red teaming to obtain diverse PHI-labeled discharge letters. In our experiments, we demonstrate this superior content diversity over existing public datasets. Further, we show SPARCLING's downstream utility as a benchmark to uncover failure modes of de-identifiers in the German clinical domain. Finally, our experiments show that training on our synthetic data can yield more capable de-identifiers than training on the scarce public data alone. Our code is available at https://github.com/Matt00n/phi_discharge_letter_generator.
Recursive self-improvement offers a promising path toward AI agents that improve through experience by refining their own harnesses without retraining their underlying large language models (LLMs). Making this self-improvement process reliable, however, requires more than generating new updates: agents must decide which changes are worth keeping. A greedy gate that accepts updates based solely on higher validation scores can mistake an observed gain for genuine progress and lead to performance degeneration, whereas overly conservative statistical gates may reject beneficial updates and hinder performance improvement. We introduce Trustworthy Recursive Updates with Statistical Testing (TRUST), a framework that adapts an online multiple hypothesis testing procedure to control the rate of false evolution across rounds, with an LLM analyzing historical context to help determine how much evidence is required to update the incumbent. Under interpretable assumptions, we prove that TRUST limits the proportion of accepted updates that result in unacceptable performance loss. We further derive a lower bound on the final performance gain. Across the benchmarks HotpotQA, DROP, and APPS with the models DeepSeek-V4-Flash and Qwen-3.8-Flash, TRUST reduces the average ratio of accepted inferior updates from the greedy gate's 45.59% to 2.78%, while achieving higher final scores than the state-of-the-art statistical gate baselines for recursive self-improvement. These findings demonstrate that TRUST strikes a balance between statistical reliability and performance gain in recursive self-improvement.
We introduce a proximal preference optimization framework that unifies the implicit reward formulation of Direct Preference Optimization (DPO) with the trust-region stability of Proximal Policy Optimization (PPO). Our framework yields two complementary algorithms. P$^2$O (Proximal Preference Optimization) introduces a clipped importance ratio directly into the DPO preference margin, producing a reward-model-free algorithm that supports safe dataset reuse, bounded policy divergence, and monotonic preference improvement guarantees. PKTO (Proximal Kahneman-Tversky Optimization) extends P$^2$O by replacing the DPO margin with KTO's prospect-theoretic utility margins, adding a self-regulating adaptive baseline that tightens the effective trust region as the policy drifts, and an asymmetric weighting that handles imbalanced or asymmetric preference datasets. We provide a unified theoretical analysis, including Kullback Leibler (KL) trust-region bounds, a policy-improvement lower bound, and convergence guarantees under multi-epoch proximal updates. Empirically, P$^2$O and PKTO match or exceed DPO and KTO in alignment quality while maintaining better stability under repeated dataset reuse across multiple LLMs on three preference benchmarks. The code base is available here.