arXiv:2601.07958v1 Announce Type: new Abstract: Speaker-specific anti-spoofing and synthesis-source tracing are central challenges in audio anti-spoofing. Progress has been hampered by the lack of datasets that systematically vary model architectures, synthesis pipelines, and generative parameters. To address this gap, we introduce LJ-Spoof, a speaker-specific, generatively diverse corpus that systematically varies prosody, vocoders, generative hyperparameters, bona fide prompt sources, training regimes, and neural post-processing. The corpus spans one speakers-including studio-quality recordings-30 TTS families, 500 generatively variant subsets, 10 bona […]

Ver mais

In this tutorial, we focus on building a transparent and measurable evaluation pipeline for large language model applications using TruLens. Rather than treating LLMs as black boxes, we instrument each stage of an application so that inputs, intermediate steps, and outputs are captured as structured traces. We then attach feedback functions that quantitatively evaluate model behavior along dimensions such as relevance, grounding, and contextual alignment. By running multiple application variants under the same evaluation setup, we show how […]

Ver mais

arXiv:2602.09055v1 Announce Type: new Abstract: This paper considers the scattering of a time-harmonic acoustic plane wave by an elastic body with an unbounded periodic surface. The original problem can be confined to the analysis of the fields in one periodic cell. With the help of the perfectly matched layer (PML) technique, we can truncate the unbounded physical domain into a bounded computational domain. By respectively constructing the equivalent transparent boundary conditions of acoustic and elastic waves simultaneously, the […]

Ver mais

I have been working on some research over the last months. I am fairly certain I have quality data and findings but as an unaffiliated researcher I often lack critical feedback. At least in my setup the refinement operation(applied additively with tanh values) is almost completely contractive along the direction of the base read. This is revealed to be necessary and the model collapses under ablation of the parallel portion. Below I have provided a link to the […]

Ver mais

For those who have been following this project, you may recall FlashLM v3, then v4 “Bolt”, and v5.2 “Nova-Ignition”. I am pleased to announce that FlashLM v5 “Thunderbolt” is now complete. Results Metric Value Final PPL 1.36 Final BPC 0.44 Parameters 29.7M (26.5M ternary) Training Time ~40 hours Hardware AMD Ryzen 7950X3D FlashLM v5 achieves a validation perplexity of 1.36, which beats the TinyStories-1M baseline (PPL 1.59). This represents the first instance of a CPU-trained model beating this […]

Ver mais

How enterprise AI is finally graduating from prototype theater to mission-critical systems engineering — and why your architecture determines everything The industry has reached an inflection point that most practitioners haven’t fully internalized yet. After spending the better part of 2024 watching organizations burn through millions on AI “pilot programs” that never left the sandbox, we’re now witnessing something fundamentally different: the wholesale transformation of AI agents from conversational toys into production-grade infrastructure components that need to meet the same […]

Ver mais

Recent reinforcement learning (RL) methods improve LLM reasoning by optimizing discrete Chain-of-Thought (CoT) generation; however, exploration in token space often suffers from diversity collapse as policy entropy decreases due to mode elicitation behavior in discrete RL. To mitigate this issue, we propose Latent Diffusion Reasoning with Reinforcement Learning (LaDi-RL), a framework that conducts exploration directly in a continuous latent space, where latent variables encode semantic-level reasoning trajectories. By modeling exploration via guided diffusion, multi-step denoising distributes stochasticity and […]

Ver mais

arXiv:2602.04992v1 Announce Type: new Abstract: Urban searches demand rapid, defensible decisions and sustained physical effort under high cognitive and situational load. Incident commanders must plan, coordinate, and document time-critical operations, while field searchers execute evolving tasks in uncertain environments. With recent advances in technology, ground-robot fleets paired with computer-vision-based situational awareness and LLM-powered interfaces offer the potential to ease these operational burdens. However, no dedicated studies have examined how public safety professionals perceive such technologies or envision their […]

Ver mais