In briefHigh-resolution land-use mapping and synthetic data augmentation in Kathmandu Valley
SummaryThe authors build a hand-annotated six-class land-use dataset of 408 high-resolution RGB scenes covering 202 km² of Kathmandu Valley. They find that random data splitting overstates performance due to spatial leakage, and a strict sequence-block split lowers results by up to 0.095 mIoU. A modified SegFormer, adding only 1.9% more parameters, improves mean IoU on every test fold, mainly for rare thin-structure classes. Augmenting training with diffusion-generated image-label pairs degraded accuracy across all architectures.
In briefInference for Reasoning Models: Lessons from Evaluation and Serving
SummaryThis book examines how inference settings, such as generation length limits, answer extraction, scoring, and quantization, shape the evaluation results and deployed behavior of language models. It notes that changing only the maximum response length moved one math benchmark's accuracy from 0.08% to 93.18%. The book covers output variation from batching, attention backends, and caching, along with serving performance, structured outputs, tool calls, reasoning budgets, and long-context tasks. It concludes with practical guidance on recording inference runs, checking evaluations, estimating capacity, and comparing systems fairly.
In briefWhat Broke & Why: Practical Lessons from Reinforcement Learning Post-Training
SummaryThis book offers a practical guide to reinforcement learning post-training for language models, including RLVR and GRPO. It follows five training programs covering math reasoning, search agents, training stability, mixture-of-experts models, and software engineering tasks. For each, it describes the problems encountered, the changes made, and the lessons learned. The book also serves as a reference on reward design, evaluation, debugging, and reproducibility.
In briefFactor Rank: a canonical algebraic generalization of tree decision diagrams
SummaryThe paper introduces algebraic tree decision diagrams (ATDDs), which replace the subfunction count in tree decision diagrams (TDDs) with the rank of the cut matrix, generalizing TDDs. For each Boolean function and vtree there is exactly one ATDD, and its canonicity can be checked in polynomial time from the tensors alone. The authors also show that vtree changes can be performed in output-polynomial time, and they formalize core results in Lean with a prototype compiler.
WhyThe authors generalize TDDs by replacing subfunction counts with cut-matrix ranks, give a canonical form over arbitrary vtrees, and report a Lean formalization.
In briefA Gamified Framework for Conflict-Driven Urban Design Using Multi-Agent Learning
SummaryThe paper models urban design conflicts as a non-zero-sum digital board game, where stakeholder agents allocate scarce resources and balance individual payoffs against shared development and resilience goals. Multi-agent reinforcement learning is used to train policies, which are tested on 4 x 4, 8 x 8, and 12 x 12 boards, the last initialized from historical maps of Tatsumi, Tokyo. On the 12 x 12 setting, trained agents average 452.1 points versus 357.3 for a random baseline over 10,000 runs. The learned strategies improve adaptability, sustainability, flood resistance, and greenery while reducing vitality and density.
In briefA simple test-time refinement method for turning plots into code, using visual-code diagnostics
SummaryThis paper studies plot-to-code generation, which produces program code that reproduces a given plot. The authors propose a Visual-Code Diagnostics framework that groups generation errors into four coarse types and trains lightweight discriminators to recognize and locate them from rendered images and code. The code is then iteratively refined based on this feedback. Experiments report consistent quality gains across diverse generators and clear improvements over strong test-time scaling baselines.
In briefLearning biological hierarchies in single-cell foundation models
SummaryThis paper introduces scOntoFM, a biologically informed single-cell foundation model that embeds the hierarchical cell ontology into representation learning. It uses Lowest Common Ancestor distances from the ontology graph together with efficient offline triplet sampling, so embeddings preserve local neighborhoods while reflecting the ontology's global structure. The authors report consistent performance gains across cell- and gene-level benchmarks, with especially strong improvements in zero-shot settings.
In briefFirepool: A system that keeps GPU pods warm for AI agent and RL jobs at LinkedIn
SummaryThis paper describes Firepool, the sandbox and execution broker LinkedIn uses for agentic and reinforcement-learning workloads. It makes GPU pods interchangeable by using shared base images and binding each job at start, and it keeps pods warm by reusing them only for the same caller while capping GPU use with queue-driven budgets. In a sampled production peak, 94.7% of 3,206 pod requests were served from warm or standby capacity. Benchmarks report agent and RL loops finishing 13x and 10.8x faster than with a fresh pod per job.
In briefCan Multimodal Models Reason Visually? A Study on Turkish Visual Reasoning
SummaryThis paper examines how vision-language models perform on high-level tasks such as pattern completion, abstract logic, and relational reasoning, using a multilingual evaluation framework that includes Turkish datasets. The authors compare different data mixing strategies and supervised fine-tuning methods, and use an LLM-as-a-Judge approach to assess reasoning quality beyond final answer accuracy. They find that baseline models often reason more stably than fine-tuned ones, and that fine-tuning can cause model collapse, such as endless text loops or empty outputs. The study concludes that current vision-language models still struggle with abstract visual reasoning.
Berke Bunyamin Sule, Yavuz Selim Aygan, Ahmed Zeer et al.
In briefAutoC2V: Training LLMs with synthesis feedback to generate better Verilog from C++
SummaryThis paper presents AutoC2V, a framework that trains large language models to turn synthesizable C++ kernels into Verilog RTL, using data built from real synthesis results. The models are aligned through supervised fine-tuning and Group Relative Policy Optimization. A QoR Prediction Reward Model estimates latency and resource metrics to provide low-cost reward signals in place of repeated synthesis runs. The authors report higher RTL functional pass rates and better latency and resource results than the base model, several HLS optimization methods, and several commercial models.
In briefLearning When to Think: Adaptive Internal Computation for Reinforcement Learning
SummaryThis paper introduces Adaptive Internal Computation (AIC), an iterative reinforcement learning policy that learns when to stop refining its latent representation and act, so computation varies by state. With at most eight refinement steps per decision, AIC averages 3.3 steps across twelve benchmark tasks. On four control and robotics tasks, it reduces per-decision latency by 31 to 44 percent compared with a fixed eight-step policy, and its mean reward beats the best fixed-depth policy on seven tasks while trailing on four.
In briefAmodalBox: Automatic 3D Box Labeling When Objects Are Partly Hidden
SummaryThis paper presents AmodalBox, an automated framework for annotating 3D bounding boxes of objects that are only partially visible to LiDAR in sparse driving scenes. It uses 3D assets generated from images as amodal geometric priors, aligning them with the partial LiDAR point clouds to estimate each object's full extent. Detectors trained on AmodalBox annotations improve over strong baselines by 14.1 mAP on nuScenes and by 12.45 AP at IoU 0.5 for vehicles on the Waymo Open Dataset.
In briefSelective cooperative perception using vision-language models to reason about blind spots and communicate through dialogue
SummaryThis paper presents a framework driven by vision-language models that reasons about blind spots and shares perception through dialogue between connected vehicles. It combines object detection, vision-language reasoning, and a geometric module, and performs cooperative perception only in safety-critical scenarios to reduce resource use. Using real-world data from connected vehicles in urban traffic, the authors report over 80% accuracy in detecting driving-critical blind spots and over 90% success in dialogue-driven perception sharing.
In briefATHENA: Hypergraph Learning for Classifying Autism in Toddlers from Eye-Tracking Data
SummaryThis paper proposes ATHENA, a model that classifies autism spectrum disorder (ASD) in toddlers using eye-tracking data, aiming for early and precise detection. It combines FMMix, a feature-level augmentation technique, with an adaptive hypergraph framework that captures complex relationships among features. On two gaze-coordinate datasets, the model reports AUC scores of 99.80% and 99.86%, and accuracies of 98.45% and 98.56%.
Quang-Anh N.D., Duc Pham Minh, Thao Pham Phuong et al.
In briefSTAR-VF: A Method That Condenses Video Data for Efficient Video Fusion Training
SummaryThe paper proposes STAR-VF, a framework that compresses an original video training set into a compact set of learnable spatio-temporal representations, which are decoded into synthetic video clips. Training uses feature-level supervision from a frozen proxy model, so ground-truth fused videos are not required. The authors report that with about 1/16 of the original temporal training set, STAR-VF achieves fusion quality comparable to models trained on the full data across infrared-visible, multi-exposure, multi-focus, and medical video fusion benchmarks.
In briefCompAIssionate Other: A study of how LLMs assume identity explains distress
SummaryThis paper examines how large language models interpret psychological distress and suicidal ideation when users disclose their social identities. Using paired counterfactual interventions, the authors evaluate six LLMs across two datasets, 16 identity and control conditions, and two prompting protocols. The findings show that models often turn identity disclosures into explanations that attribute distress to presumed identity-related adversity, especially when marginalized identities are mentioned, and that these assumptions appear within otherwise empathetic responses. The authors frame this as a harm they call compassionate othering and argue that distress should be interpreted based on users' own disclosures rather than unsupported identity-linked assumptions.
Drishti Goel, Ethan M. Nguyen, Violeta J. Rodriguez et al.
In briefWARP-JEPA: Faster Planning for Frozen Hierarchical Visual Control Models
SummaryThe paper addresses hierarchical joint-embedding predictive architectures (H-JEPA), which rely on slow test-time gradient descent for planning in visual control. It proposes WARP-JEPA, which replaces online optimization with learned multi-candidate trajectory proposers while keeping the world models frozen. Across 30 frozen H-JEPA models on four benchmarks, it reports 9.6 to 18.8 times faster episode inference on long-horizon tasks and up to 33.3 percentage points higher task success.
Junjie Li, Zhaojiacheng Zhou, Kaixiang Wang et al.
In briefGeneralization in bilevel least-squares learning with overparameterized neural networks
SummaryThis paper studies how well bilevel optimization generalizes when both the inner and outer problems are least-squares regression. The authors estimate the hypergradient through implicit differentiation in function space and derive stationarity bounds for inexact gradient methods in terms of the uniform generalization errors of the two regression functions. For overparameterized two-layer ReLU networks trained by gradient flow with early stopping, they obtain explicit finite-sample bounds in the Neural Tangent Kernel regime. Experiments on synthetic data illustrate the effect of sample size and show an initialization-dependent implicit bias beyond that regime.
Fares El Khoury, Edouard Pauwels, Samuel Vaiter et al.
In briefEvaluating Jev on Arabic: Comparing Decision Models and LLMs Across Arabic Tasks and Dialects
SummaryThis paper evaluates 67 systems, including 47 decision models and 20 large language models, on 43 Arabic tasks across 9 task families using a shared zero-shot protocol. Jev 1.14.0 scores 69.24 on the mean primary metric, while Gemma-4-31B scores highest at 71.81. Removing English controls or repeated items leaves the system ranking unchanged. The authors also compare Modern Standard Arabic with dialects and conclude that aggregate scores should be read alongside task-family results.
Hossam Elsafty, Tobias Deu{\ss}er, Lorenz Sparrenberg et al.
In briefImproving Turkish Math Reasoning in Language Models Through Search and Model Merging
SummaryThis paper improves mathematical reasoning in Turkish large language models by combining LoRA fine-tuning, weight-space model merging, and search-based inference. Using Turkish-Llama 8B-DPO trained on Turkish translations of MetaMath, NuminaMath, and TigerLab, the authors find that data quality matters more than data scale. Linear merging gives the largest gains, reaching GSM8K 0.63, while search methods (MCTS and Tree of Thoughts) raise LLM-judged reasoning quality scores.
Ahmed Zeer, Osama Shbib, Muzaffer Kaan Yuce et al.
In briefRTLREX: LLM-guided framework for exact high-level hardware design recovery
SummaryWhen the original register-transfer-level (RTL) design is unavailable and only gate-level netlists or input-output descriptions remain, RTLREX uses a large language model to coordinate domain-specific tools for function recovery, RTL generation, independent verification, and cost-guided rewriting. Independent verification establishes that the recovered RTL is equivalent to the original circuit for all input combinations. On 30 benchmarks from ICCAD 2022 CAD Contest Problem A, the framework recovers equivalent high-level RTL for all benchmarks, surpassing the state-of-the-art method WolFEx in recovery coverage.
In briefAdapting Large Language Models via Multi-Task Learning for Mental Health and Emotion Detection
SummaryThis study fine-tunes Llama 3.1 8B Instruct using sequential multi-task learning with LoRA, 4-bit quantization, and parameter-efficient fine-tuning for mental health problem, emotion, and sentiment detection. Using 4,243 English social media posts, it compares two task orders and finds that task ordering significantly affects knowledge transfer and interference. The mental health problem-first order performs best on mental health detection, while the emotion-first order performs better on emotion and sentiment analysis. The paper also notes challenges of joint learning, such as task interference and generative format collapse.
In briefZeroth-order Riemannian optimization on low-rank matrix manifolds for memory-efficient LLM fine-tuning
SummaryThis paper proposes constraining low-rank weight updates to a fixed-rank matrix manifold while keeping pretrained weights frozen, so updates stay rank-constrained during fine-tuning. It builds a Riemannian zeroth-order gradient estimator and introduces ZO-RGD and its smoothed variant ZO-RGD-S. Experiments on LLM fine-tuning tasks show the methods are competitive with strong forward-only baselines, with gains in several settings.
In briefGenerating diverse, verified synthetic clinical notes with built-in labels
SummaryThis paper presents SPARCLING, an LLM-based pipeline that generates synthetic discharge letters with character-exact annotations of protected health information (PHI) spans. The pipeline combines structured attribute sampling, template generation, and red teaming to produce diverse PHI-labeled letters. Experiments show higher content diversity than existing public datasets, and the data can serve as a benchmark that reveals failure modes of de-identification tools in German clinical text. Models trained on the synthetic data also outperform those trained on the scarce public data alone.
In briefMaking AI self-improvement more reliable through statistical checks
SummaryThis paper introduces TRUST, a framework that lets AI agents improve by refining their own harnesses without retraining the underlying language models. It uses an online multiple hypothesis testing procedure to control how often harmful updates are accepted, with an LLM analyzing history to set how much evidence an update needs. On HotpotQA, DROP, and APPS, the share of accepted inferior updates fell from 45.59% under a greedy gate to 2.78%, with higher final scores than prior statistical gates.
In briefProximal Preference Optimization: A stable, reward-model-free method for aligning AI with human preferences
SummaryThis paper proposes a framework that combines the implicit reward approach of Direct Preference Optimization (DPO) with the trust-region stability of Proximal Policy Optimization (PPO). It introduces P²O, which adds a clipped importance ratio to the DPO margin and needs no reward model, and PKTO, which builds on P²O using KTO's prospect-theoretic utility margins, an adaptive baseline and asymmetric weighting. The authors provide theoretical analysis covering KL trust-region bounds, a policy-improvement lower bound and convergence guarantees. Experiments on several LLMs and three preference benchmarks report alignment quality matching or exceeding DPO and KTO, with better stability under repeated dataset reuse.
In briefEEGFaceSem: EEG benchmark pairing brain signals with generative face latents
SummaryThis paper introduces EEGFaceSem, a large-scale dataset pairing EEG recordings with photorealistic face images synthesized by a generative model, along with a latent vector for each stimulus. Thirty participants attended to semantic features such as smiling or hair color during rapid serial visual presentation, yielding 67,200 time-locked EEG recordings. The authors provide benchmarks for single-trial semantic classification and cross-subject voting, and demonstrate two applications: brain-supervised image editing and brain-conditioned image generation.
Jun Ma, Michiel M. Spap\'e, Keith M. Davis III et al.
In briefSame measurements, different velocity risk: complementary camera readouts under timing uncertainty
SummaryThe paper studies how two cameras with rolling shutters estimate velocity when exposure timing is uncertain. Repeated readouts have a worst-case velocity risk that shrinks only on the order of the noise level, while complementary readouts achieve minimax risk on the order of the noise variance. Mounting configurations where both designs are regular are identified, and a risk floor caused by metric uncertainty is quantified.
In briefUniversal approximation proof for a graph neural network interatomic potential
SummaryThis paper proves that the bilinear-readout equivariant message-passing network (BREMP), a reference architecture for equivariant graph neural network interatomic potentials, can universally approximate atomic energies. It constructs a one-layer feature that determines the local atomic environment up to permutations of same-type atoms, and shows the readout approximates any continuous invariant function of these features. The authors also state that DPA4, an architecture in production use, inherits this result by approximating the BREMP layers constructed in the proof.
