NVIDIA Research

This work training a LALM with reinforcement learning to reason over its own speech output generates a draft speech as a form of audio-token reasoning, critiques its own generation by reflecting on the acoustic realization in text, and produces a refined version conditioned on both the first-pass speech and the critique.

This work introduces PosteriorBench, a benchmark for evaluating the distributional accuracy of generative inverse solvers, and reveals substantial distribution-matching gaps across current solvers, while showing that neural operators improve resolution robustness, and guidance weights and generation noise are key to posterior-variance calibration.

FloatLib is the first Lean library to unify IEEE binary and decimal arithmetic, arbitrary-width posits, P3109, and user-defined formats and rounding rules behind interchangeable certified software backends.

Industry-standard GPU memory sanitizers harbor major blind spots across core memory spaces, finally quantifiable through 149 targeted CUDA stress tests.

EvoSafeHarness is a safety-specific optimization framework that synthesizes a deployable harness for a frozen model in a target domain, guided by model behavior, domain specifications, and fresh-context adversarial review to reject benchmark-specific rules.

Fine-tuned robotic policies can be biased towards certain instruction factors, but a new bias-aware data strategy can significantly enhance their performance with fewer demonstrations.

Gauntlet outperformed human reviewers in technical critique of computer architecture papers, revealing that LLMs can achieve significant analytical depth through structured multi-agent collaboration.

Current VLMs struggle with specialized domains, failing to adapt effectively in both zero-shot and ICL scenarios, revealing critical gaps in their spatio-temporal reasoning abilities.

ASR models can exhibit drastically different performance depending on user preferences, revealing hidden quality disparities in traditional benchmarks.

Current multimodal conversational models miss critical emotional cues, revealing a significant gap in their ability to engage in nuanced human-like dialogue.

LLMs can pass Taiwanese lawyer qualification exams but struggle with precise legal citations, revealing critical gaps in their legal reasoning capabilities.

Self-reports in LLMs show selective coherence, revealing that traditional personality frameworks like the Big 5 may be inadequate for predicting behavior in deployment scenarios.

Current vision-speech agents are surprisingly bad at mimicking the subtle, real-time audio-visual cues that make human conversation feel natural.

Stop wasting compute on redundant code generation attempts: CPPO boosts pass@$K$ by explicitly coordinating exploration across diverse algorithmic strategies.

LLMs can discover interpretable and effective criteria for assessing communication quality, outperforming traditional content-based methods in predicting real-world impact.

Existing robotic methods falter in tackling fundamental physical reasoning challenges, as evidenced by KinDER's rigorous benchmark evaluation.

Systematic biases in LALMs can be triggered by subtle cues like gender and accent, revealing a complex landscape of fairness that traditional benchmarks miss.

Scaling up LLMs doesn't uniformly improve context handling; instead, it paradoxically amplifies the tendency to copy irrelevant tokens while simultaneously improving resistance to misinformation.

RoboLab exposes critical performance gaps in leading robotic models, revealing that high-fidelity simulations can better assess generalization than traditional benchmarks.

A hybrid cuVSLAM-based visual SLAM system achieves superior mapping accuracy in real-world logistics environments, outperforming other VO/VSLAM approaches.

Training generalist robots just got a whole lot easier: RoboCasa365 offers a massive, diverse, and reproducible benchmark for household mobile manipulation.

Forget simulated manipulation—ManipulationNet offers a global infrastructure for benchmarking robots in the real world, complete with standardized hardware and software, to finally measure progress toward general manipulation.

Current NVS evaluation metrics are misleading, so this paper introduces a task-aware framework using Zero123 features that actually aligns with human perception of quality and faithfulness.