Infrastructure & Hardware B

Model merging integrates the capabilities of several models fine-tuned from the same pretrained checkpoint into one, enabling multi-task learning. This work proposes Essential Subspace Merging, which extracts and merges each task's essential subspace to reduce interference and preserve multi-task performance.

IEEE Spectrum examines how musicians can be compensated when their music is used to train AI, covering attribution and payment for training-data use. This summary is title-based as the excerpt was blocked by a cookie/query-string wall and not retrieved; the specific mechanisms are per the article and unverified independently.

Identifying the lowest-energy surface-adsorbate configuration is critical for modeling heterogeneous catalysis, but exhaustive ab initio exploration is prohibitive. AdsMind is a physics-grounded multi-agent system that self-corrects to efficiently discover adsorption configurations.

The prevailing intuition for conceptual drift in long-horizon LLM collaboration is to trade more formal constraints for more reliable outputs. Across 391 consecutive sessions, this work studies semantic space control and 'index sickness' elimination in workflows written and managed by AI.

LLMs are increasingly used as judges for scalable evaluation, yet LLM-as-a-Judge systems show systematic biases decoupled from semantic quality, notably verbosity bias. This work uses positive-unlabeled learning to quantify and audit LLM evaluation, helping detect and correct such biases.

The mismatch between continuous acoustic signals and discrete event-driven processing is a fundamental bottleneck for neuromorphic speech processing. Rather than fixed spike encoders, this work proposes adaptive speech-to-spike encoding for spiking neural networks, improving downstream performance.

Pretraining LLMs typically demands large-scale infrastructure with tightly coupled accelerators. As model and data scale grow, FoMoE proposes a federation of Mixture-of-Experts that avoids replicating the full model across devices, breaking the full-replica barrier and easing infrastructure constraints.

Reinforcement learning post-training of Diffusion Transformers is prohibitively expensive, needing thousands of high-end GPUs. Spotlight synergizes seed exploration with cheap, preemptible spot GPUs to substantially reduce the cost of DiT RL post-training.

Model merging effectively composes the capabilities of a multilingual model and a reasoning model, achieving promising generalization on multilingual reasoning by aligning their feature spaces. This work introduces steerable model merging to control the composition and further boost multilingual reasoning.

Agents are increasingly deployed in document-intensive workflows where sensitive private information is routine input—e.g., booking a flight needs passport numbers. TRAP is a benchmark evaluating agents on both task completion and resistance to active privacy-extraction attempts.

Idioms resist literal cross-lingual mapping because they are non-compositional. G-IdiomAlign anchors each idiom to an English Wiktionary gloss and adds a high-confidence reference alignment set. Two protocols (multiple-choice idiom equivalence and gloss-contrastive generation) isolate the effect of explicit glosses.

Production LLM agents increasingly depend on real-time search but get locked into vendor-specific grounding. This work decouples search from reasoning with a vendor-agnostic grounding architecture, letting search backends be swapped while preserving reasoning quality.

PLCopen XML defines encodings for IEC 61131-3 Ladder Diagrams. Graph-ESBMC-PLC applies SMT-based model checking to formally verify graphical PLCopen XML Ladder Diagram programs, supporting correctness checking of industrial control software.

Sequence labelling is a core NLP task. This work proposes an approximate structured diffusion approach that models label dependencies while keeping sequence labelling efficient.

Turkish is agglutinative, with meaning carried by morphemes that subword tokenizers fail to capture. Morpheus is a morphology-aware neural tokenizer and word embedder designed to improve Turkish language processing.

Cohere details how it ensures every tenant gets a fair share of compute in LLM serving, tackling the 'noisy neighbour' problem where one user monopolizes resources. The design allocates capacity fairly across tenants to deliver stable, predictable multi-tenant performance.

NVIDIA introduced NVIDIA XR AI, a framework for developers to build AI agents for AR glasses and wearable XR devices. It targets the gap between ready hardware and the work of integrating live, real-time AI experiences. Capabilities are per NVIDIA's own announcement; third-party verification pending.

NVIDIA's developer blog walks through how to build your own transaction foundation model aimed at financial-intelligence use cases such as fraud detection and risk analysis. Specifics and claimed benefits come from NVIDIA's own post; independent verification is pending, as the raw excerpt was unavailable and this is summarized from the title and source.

ITmedia surveys 'physical AI'—a strategic focus area for Japan's government—through the lens of autonomous driving. The article reviews how advances in generative AI are reshaping the competition and compares the latest moves and differing approaches of Tesla, Waymo and NVIDIA.

Reliable physics simulation needs Young's modulus, Poisson's ratio and density, which most 3D assets lack. AdaVoMP predicts dense, spatially varying values of these properties for input 3D objects in a way invariant to resolution across representations.

Studying finite-horizon queue peaks in generalized switches, where many queues share constrained service resources, the paper derives laws under a uniform interior-slack load condition showing logarithmic scaling of peaks after geometric thresholds.

NVIDIA announced the ACE Game Agent SDK and Unreal Engine 5 plugins for developers to build on-device AI companions—AI agents that run locally on the device rather than in the cloud—for in-game characters. The export raw_excerpt was blocked (cookie/query string data), so this is summarized neutrally from the title and the NVIDIA developer blog framing; specific figures and performance claims are unverified.

The paper formalizes 'cognitive atrophy,' a process-level behavioural measure of AI-mediated mental-health support, capturing whether interactions help users keep reflecting, coping, and deciding, a dimension distinct from safety and static response quality.

Conversational AI localizes responses using user metadata, yet the regional biases this hidden context introduces remain poorly understood. The paper analyzes the unintended effects of geographic conditioning on large language model outputs.

Ternary Mamba applies grouped quantization-aware training to Mamba state space models with ternary (W1.58) weights and 16-bit activations, targeting efficient low-bit training and inference of sequence models while preserving accuracy.

RAG grounds model outputs in external evidence, but its dominant evaluations and defaults are oriented toward factual question answering. HistoRAG embeds historical methodology into retrieval-augmented generation through critical technical practice for interpretive historical studies.

An NVIDIA technical post explains techniques for optimizing transformer-based models during low-precision training. The export raw_excerpt was blocked (cookie/query data), so this summary is based only on the title and source; specific methods and figures are unverified.

To uncover causal dependencies among variables, the paper proposes TSCD, a tensor-based second-order causal discovery algorithm whose input is a tensor formed from covariance matrices of observational and interventional data, assuming linear structural equations.

The paper proposes a multi-agent framework for healthcare that mitigates premature diagnostic handoff and silent clinical hallucinations, replacing LLM-as-a-judge routing with deterministic orchestration constraints and adding two safety mechanisms.

Decision-making with deep time-series forecasting needs not just accurate predictions but actionable insight, which current architectures lack. ConTex reformulates counterfactual generation to indicate how present conditions must change to shift a predicted outcome toward a desired future.