Skip to main content
Search

Nvidia (United States)

18 related articles

Subscribe to updates
  1. arXiv

    Multi-talker transcription stops decoding each speaker alone: Pandemonium lets speakers see each other and reaches 14.7 average cpWER on meeting benchmarks

    Target-speaker recognition for multi-talker meetings previously decoded each speaker independently and was confined to 30-second windows. Pandemonium moves the diarization conditioning onto the longform Parakeet backbone, which decodes a whole conversation in one pass, and adds cross-speaker attention plus a mask refiner so that the speakers of one conversation become visible to each other inside the encoder. Trained on synthetic, pseudo-labeled and real data, it reaches state-of-the-art results across several meeting benchmarks, with an average cpWER of 19.2 under predicted diarization.

  2. Finite-element solvers become differentiable graph programs: one H200 runs a step up to 181 times faster than 32 CPU cores

    A finite-element solver for conjugate heat transfer and buoyant flow is recast as fixed-weight graph programs inside an automatic-differentiation framework, with assembled operators and the algebraic-multigrid cycle executed as sparse graph operations and nothing trained. With discretization, mesh, time step and tolerances held fixed, the same problem runs up to 181 times faster per step on one H200 than on 32 CPU cores, and the pressure solve alone 18.7 times faster. Reverse-mode gradients match central differences, and implicit differentiation holds one gradient near 4.9 forward solves.

  3. arXiv

    Simulation cases stop being re-curated per model: whole-case loads run 88x faster than VTK, and overlapped preprocessing costs 2.1% of a step

    PhysicsNeMo-Mesh stores a simulation case as a GPU-native mesh with named boundaries and prescribed condition values, moving preprocessing from dataset curation to load time at each training step. On one x86 machine with local NVMe, a whole-case load is 88 times faster than reading VTK files; on one GB200, mesh operations run 6.2 to 1068 times faster than threaded VTK, and overlapped with training, preprocessing costs 2.1% of throughput for a 6.0M-parameter Transolver. In a synthetic heat-conduction study where geometry and wall conditions both vary, per-boundary conditions cut error from the geometry-only floor to 0.038–0.066.

  4. arXiv

    Speech LLMs stop pooling at a fixed frame rate: a learned boundary policy cuts English audio tokens by about a quarter

    ASSET trains a speech encoder's segment boundaries with reinforcement learning, reaching recognition close to a transcript-assisted character-boundary oracle on LibriSpeech with roughly 25% fewer audio tokens.

  5. arXiv

    SSM decoding stops writing state every step: caching recent inputs halves state traffic and matches vLLM speculation

    State-space model decoding previously read and wrote the full recurrent state at every step, making state traffic the memory-bandwidth bottleneck and leaving speculative decoding without cheap rollback. ReplaySSM represents the state as a checkpoint plus a buffer of recent inputs, writes the checkpoint only when the buffer fills, and computes outputs directly on most steps. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it speeds up standard decoding over a tuned baseline at serving batch size 256, reaches about the throughput of vLLM's speculative decoding on two 120B-class models, and separate state precisions reduce output KL divergence and downstream accuracy losses.

  6. arXiv

    One bitstream diffusion recipe spans images, speech and proteins: 3.5% word error in joint speech-text, but two fold-level protein figures disagree

    Continuous bitstream diffusion writes images, speech and proteins as binary codes and, with one training and sampling protocol, reaches 3.5% word error rate in joint speech-text generation, while the fold-level agreement of generated protein sequence-structure pairs is reported as 75% in the results section and 69% in the conclusion.

  7. arXiv

    Agents stop reading each other's rewards: watching peers' behavior is enough to cooperate and split returns more evenly

    Multi-agent reinforcement learning no longer needs peers' private rewards: each agent evaluates others' transitions with its own learned reward model, learns cooperation in Escape Room, Clean Up, and Commons Harvest, and often divides returns more evenly than agents that see the true rewards.

  8. arXiv

    MemCo pairs local and global memory so LLM agents gain 43.47% held-out ALFWorld accuracy and 32.10% fewer steps in unseen environments

    The authors propose MemCo, a memory-centric collaboration framework in which each agent keeps an environment-specific local memory graph while evidence-aggregated, Wilson-lower-bound-filtered transferable workflows are promoted to a shared global memory, and at decision time local and global memories are retrieved and grounded by current state, task phase, and goal; across ALFWorld, PDDL, FEVER, and ScienceWorld with multiple LLM backbones, MemCo attains the highest accuracy in 10 of 12 evaluation settings and the fewest interaction steps in 9, e.g., on Qwen3-4B it improves ALFWorld held-out layout accuracy by 43.47% and reduces steps by 32.10% relative to the runner-up baseline.

  9. arXiv

    PyRUA-Lean lifts GPT-6 Astra robot-agent success from 63.1% to 71.7% while cutting input tokens by 65%

    The authors introduce PyRUA-Lean, an interactive code-execution framework that couples feedback-driven primitive composition with selective observation, letting the agent compose classical robot primitives and learned vision-language-action (VLA) policies into Python cells that perform conditional checks and local retries and return only explicitly requested images and state feedback for replanning; across 700 simulated task instances from LIBERO-PRO, RoboTwin 2.0, and RoboCasa365, compared with a tool-calling baseline using the same GPT-6 Astra planner and the same underlying robot primitives, it raises overall success from 63.1% to 71.7% under equal LLM-call budgets and, on instances solved by both agents, uses 49% fewer LLM calls and 65% fewer input tokens.

  10. arXiv

    PLUTO turns natural-language rendezvous missions into SCP constraints, generating a trajectory for all 50 prompts with 94% passing mission-intent verification

    The work introduces PLUTO, an interactive trajectory design framework that embeds agentic AI coding tools within a sequential convex programming (SCP) architecture, mapping natural-language rendezvous mission requirements into structured mathematical constraints that pass through an auto-convexification pipeline into executable optimal control formulations; across 50 natural-language rendezvous mission prompts spanning 10 constraint types, PLUTO generated a trajectory for every prompt and 94% of the resulting trajectories satisfied both numerical constraint checks and semantic consistency with the original mission intent.