National University of Singapore

Diffusion models can learn new objectives by stochastically routing attention within their existing architecture, achieving 4x faster RL training without sacrificing generation quality.

Driving models can achieve near-perfect performance with just one denoising step by learning to predict visual motion directly, outperforming multi-step prediction models and achieving 3.29x inference speed.

Synthetic experience unlocks substantial game control improvements for LLMs, pushing success rates on complex benchmarks like MineDojo-300 from 12.63% to over 38%.

Autonomous driving VLAs can be made conflict-aware by dreaming world models that strictly follow navigation signals, revealing unsafe instruction scenarios.

A single, stage-specific objective per expert in a decoupled diffusion distillation framework enables state-of-the-art video generation with a 10,000x reduction in training cost.

A conditional energy prior explicitly guides latent space towards perturbation-specific signals, unlocking an 83.2% improvement in perturbation recognition for cellular response prediction.

Quality-aware pretraining strategies, particularly approach-aware episode conditioning, can significantly outperform plain behavioral cloning in VLA tasks, with optimal choices depending on label granularity and model architecture.

A novel bandit algorithm finds the optimal self-speculative decoding configuration for VLMs in real-time, boosting inference speed by up to 11% without retraining.

State-of-the-art coding agents exhibit critical boundary violations, with nearly all misinterpreting dependencies and over half compromising environment integrity or user inputs.

VLMs can be taught to *use* latent geometry by training them to read continuous geometric tokens, rather than just reconstructing geometry from RGB.

A novel CSC MOS structure disentangles interface and shallow trap effects in AOS transistors, revealing distinct trap regimes and their time-dependent contributions to conductance loss.

DiffusionART achieves state-of-the-art RL fine-tuning for diffusion models, reaching peak performance with over 12x fewer GPU hours and substantially higher rewards than existing methods.

Robots can now leverage diffusion models for real-time control by extracting actionable uncertainty estimates in milliseconds, bypassing costly Monte Carlo simulations.

Robots can now achieve both smooth, long-horizon motion and rapid reactivity by continuously refining action chunks with REACT, eliminating the traditional control trade-off.

Frontier LLMs can achieve substantial recursive self-improvement, closing nearly half of the performance gap on complex benchmarks by jointly optimizing data, training, and execution harnesses within a unified, agent-native environment.

Medical LLMs can achieve superior QA performance by dynamically selecting specialized adaptation components based on question context and clinical tags, surpassing static parameter-efficient methods.

Adapters forget as much as they learn: a simple post-training normalization unlocks significant gains in both specialization and retained capabilities for LoRA.

A single time-series language model now masters forecasting, contextual prediction, and temporal reasoning, outperforming specialized models on key benchmarks.

Future needs, not current context, are the key to unlocking effective long-horizon video generation.

Standard RLVR observations are mathematically insufficient to isolate or fix verifier exploitation, proving that reasoning models cannot unlearn rewarded errors without external audit signals.

High-level robot planners routinely fail because they treat low-level skills as opaque black boxes; turning training-time inductive biases into explicit compositional interfaces bridges the gap to robust out-of-distribution execution.

Non-signaling correlations provide zero advantage in boosting classical Rényi capacity or beating error exponents, proving that even hypothetical supra-quantum resources cannot overcome fundamental channel simulation limits.

Agents can achieve up to 5.8% higher success rates by learning a reusable "experience tree" directly from data, bypassing LLM-generated skills and enabling true generalization.

Full-scale 3D photoacoustic imaging no longer requires trading transducer geometry for computational feasibility: replacing surface discretization with analytical geometric intersections delivers exact forward-operator modeling at 10x the speed.

This work proposes OPTED (on-policy fine-tuning for end-to-end driving) which decouples reinforcement learning from the post-training of the end-to-end policy: a privileged teacher is trained using RL on vectorized inputs (HD-map and bounding boxes) which provides supervision to the pre-trained student during closed-loop post-training.

This work presents a preprocessing technique that improves the efficiency of ZkUnsat without introducing additional leakage, and normalizes the proof so that each derived clause is justified by a resolution chain of fixed public length k, which eliminates chain-length leakage and reduces prover memory usage.

Benchmark-average rankings mask massive sample-level disagreement among vision pruning techniques: dynamically routing inputs across existing pruning methods yields a 26.9% relative accuracy jump over any single fixed strategy.

Scaling generalized zero-shot recognition to massive vocabularies does not require re-architecting backbones: treating seen-versus-unseen gating as a downstream Monte Carlo calibration problem boosts unseen accuracy by over 20% in just 15 dimensions.

Interactive communication yields zero minimax advantage in 1-bit distributed mean estimation: purely non-adaptive queries match the adaptive rate for any heavy-tailed moment condition $k > 1$.

Achieving an 11% boost in energy efficiency for edge AI applications by smartly balancing throughput and latency through hierarchical operator parallelism.

Iterative problem-solving reveals that LLMs, while less accurate, can generate insightful solutions that illuminate failure modes in student coding attempts.

Transforming scientific papers into editable vector figures, \textsc{FigTree} makes figure generation and refinement as seamless as writing.

RailSyn achieves up to 4.9 points improvement in RFOD detection accuracy by intelligently generating synthetic data that addresses specific task deficiencies.

As human oversight wanes, LRMs could autonomously evolve, but this shift introduces significant risks like reward hacking and feedback drift.

Reducio slashes memory requirements for serverless cloud deployments while enhancing privacy, all without the need for costly infrastructure overhauls.

PriMD revolutionizes multimodal emotion recognition by enabling robust performance even when critical data modalities are missing.

Missing spatial constraints in CAD generation can be effectively filled using past design experiences, significantly enhancing the accuracy of text-to-CAD translation.

Quantum kernels can significantly boost fraud detection accuracy by aligning feature-map geometry with interaction-sensitive decision boundaries, outperforming traditional methods.

Achieving up to a 72.81× speedup in long-sequence inference could redefine efficiency standards for Diffusion Language Models.

Agents using ParallelWorld can efficiently evaluate multiple future trajectories, leading to superior decision-making in complex environments.

RecVerse outperforms traditional simulators by maintaining a nuanced understanding of user intent and memory, leading to more realistic shopping behaviors.

SPACE recalibrates multivariate forecasting by leveraging current sample clouds, resulting in improved coverage efficiency that outperforms traditional methods.

Cheap auxiliary signals can provide unbiased conditional performance estimates for language models, revealing hidden strengths and weaknesses that gold labels alone miss.

MSCNet not only reconstructs missing MRI sequences but also enhances lesion fidelity, outperforming traditional methods in clinical settings.

Internal visual modality entropy can serve as a reliable self-evaluation metric for action generation in heterogeneous VLA architectures, outperforming traditional methods.

Current AI-generated videos that mislead viewers are also the most challenging for existing detection systems to identify, revealing a critical vulnerability in misinformation defenses.

Making the teacher's privileged context learnable end-to-end enables agents to evolve more efficiently, outperforming traditional methods with less than 30% of their rollout budget.

StreamFlow achieves a remarkable 67.73% accuracy in streaming video understanding while cutting latency and memory usage by over 50%.

SafeCA slashes jailbreak success rates by 20% while adding virtually no latency, revolutionizing defenses for text-to-video models.

Forgetting in hyperbolic continual learning is driven by semantic drift and hierarchical distortion, revealing critical insights for preserving multimodal representations.

Achieving state-of-the-art performance in both 4D reconstruction and point tracking, Uni4R leverages continuous velocity fields to model dynamics at any timestamp, breaking free from traditional limitations.

ST-Omni-R1 not only excels in sound-event recognition but also sets a new standard for spatial audio reasoning, achieving nearly double the accuracy of existing models.

CED reveals that VLMs can be trained to prioritize evidence-based reasoning over language shortcuts, leading to more reliable visual understanding.

F$^2$Agent achieves over 20% better annualized returns than existing models by dynamically capturing inter-modality dependencies and resisting market noise.

Models that ignore context may seem robust, but they can fail spectacularly when the context is actually trustworthy.

World rehearsal enables LLM agents to internalize environment dynamics, achieving superior performance without costly external interactions.

PrivDPO achieves robust LLM alignment while maintaining privacy, outperforming traditional methods in balancing privacy and utility.

Task-conditioned wrist modeling boosts robot manipulation accuracy, revealing how wrist interactions can be anticipated for better performance.

Modality Balance can be harnessed as a powerful form of privileged information, leading to substantial gains in reasoning performance for multimodal models.

The first publicly available dataset for early pregnancy fetal ultrasound screening could revolutionize automated diagnostics and standardization in prenatal care.

RAG-Stack uncovers quality-performance trade-offs in RAG systems, achieving up to 153% more effective configurations than current methods.

Hierarchical tokenization transforms bounding-box representation, leading to substantial improvements in localization accuracy and model performance across various benchmarks.

Personalization in LLM agents is more complex than previously thought, with existing benchmarks failing to capture the dynamic nature of user preferences and their impact on task execution.

Hypergraph-based failure attribution can boost LLM reasoning accuracy by efficiently pinpointing the root causes of errors, outperforming traditional methods.

SPEAR achieves a staggering 99.5% increase in click recall by aligning query rewrites with user intent, transforming e-commerce search effectiveness.

Personalized 3D avatars and LLM-driven therapists can significantly enhance emotional bonding in digital mental health interventions.

SPIRAL aligns visual representations with textual semantics, achieving near-native performance in Vision-Text Compression without external supervision.

Search effort doesn't guarantee better answers; instead, the quality of retrieved evidence is the true driver of accuracy in long-horizon search agents.

DART achieves a remarkable 75% reduction in inference cache size while enhancing associative recall in long-context sequence modeling.

Low-frequency bias in EEG models can be corrected, leading to a significant boost in performance across a wide range of tasks.

Fetch-then-Explore enables search agents to retain and revisit selected pages, significantly enhancing their accuracy and efficiency in information retrieval.

Redundant neuron encoding can thwart white-box attacks, ensuring safety without sacrificing model performance.

Switching document formats can lead to accuracy drops of over 53%, revealing a hidden vulnerability in LLM workflows that demands urgent attention.

ConMem cuts input tokens by 88.2% while boosting QA accuracy to 76.0%, revolutionizing how inspection logs are utilized in risk assessment.

Even the best vision-language models struggle with reliable evaluation of computer-using agents, but OS-Shepherd models offer a low-cost solution that matches their performance.

Incomplete modalities can lead to significant distortions in multimodal recommendations, but CaIRec effectively calibrates item representations to enhance performance and coherence.

CoRAS can significantly reduce the number of measurements needed for accurate image reconstruction while ensuring error thresholds are met, outperforming fixed-rate methods.

FleetScape reveals that spatial interaction can significantly enhance drone fleet supervision, but larger fleets may compromise situational awareness.

SkillRise achieves up to 8.5 percentage points better performance than leading methods by effectively reusing transferable skills across related tasks.

Action diversity is most valuable when a VLA model is likely to fail, leading to a 17.3% boost in success rates without unnecessary perturbations when success is probable.

Predicting human decisions requires understanding not just the physical world, but also the mental states that drive behavior—MWM makes this explicit.

Explanation quality is a critical yet overlooked dimension of LLM agent performance, with many agents generating misleading explanations that can lead to incorrect code assessments.

Functionalization can lead to chemically distinct changes in electronic structure, revealing critical insights into lithium-metal electrolyte behavior.

Proprietary MLLMs may achieve high diagnostic accuracy, but they still struggle with reliable clinical reasoning, revealing significant gaps in their practical utility.

LaP-Forensics reveals that leveraging reconstruction-based evidence can significantly enhance deepfake detection accuracy against state-of-the-art generative models.

Existing AI-generated image detectors falter dramatically, with accuracy plummeting from 91-96% to as low as 54-66% when faced with realistic manipulations.

Coding agents struggle to find the right context, with logged trajectories missing gold files in 27-35% of cases, revealing critical gaps in retrieval effectiveness.

Task transformation enables LLMs to achieve self-improvement in open-ended tasks without the biases and costs associated with human evaluators.

PrefReward reveals how an explicit user preference matrix can drastically enhance personalization in text generation, outperforming traditional methods in both quality and interpretability.

IRIS can detect model substitutions and routing dilutions in LLM gateways with unprecedented accuracy using only the output text, challenging the reliability of commercial AI services.

Hallucinations in chemical reasoning models coexist with correct answers, revealing a complex relationship that challenges our understanding of model reliability.