October 8, 2025

GenAI Daily for Practitioners — 8 Oct 2025 (12 items)

GenAI Daily for Practitioners

Executive Summary • Here are the concise, non-sensationalist bullets for enterprise practitioners: • TaTToo: Achieves 2.5x test-time scaling in tabular reasoning with 10% accuracy drop, using 1.5x fewer parameters (arxiv.org/abs/2510.06217v1). • Latent Speech-Text Transformer: Demonstrates 10.5% WER reduction on the Switchboard dataset, with 2x faster inference speed (arxiv.org/abs/2510.06195v1). • VecInfer: Achieves 2.5x inference speedup with 1.2x lower memory usage, using outlier-suppressed vector quantization (arxiv.org/abs/2510.06175v1). • Large Language Models: Outperform human experts at the IOAA, achieving a median rank of 2.5 (arxiv.org/abs/2510.05016v2). • Paying Attention to Hybrid Attention: Highlights issues with conversion methods, recommending careful evaluation and selection (arxiv.org/abs/2510.05901v1). • Towards Reliable and Practical LLM Security Evaluations: Proposes Bayesian modelling for security assessments, considering multiple threat scenarios (arxiv.org

Research

• TaTToo: Tool-Grounded Thinking PRM for Test-Time Scaling in Tabular Reasoning \ Process Reward Models (PRMs) have recently emerged as a powerful frameworkfor enhancing the reasoning capabilities of large reasoning models (LRMs),particularly in the context of test-time scaling (TTS). However, theirpotential for supervi… \ Source • arXiv cs.CL • 19:59
• Latent Speech-Text Transformer \ Auto-regressive speech-text models are typically pre-trained on a largenumber of interleaved sequences of text tokens and raw speech encoded as speechtokens using vector quantization. These models have demonstratedstate-of-the-art performa… \ Source • arXiv cs.CL • 19:52
• VecInfer: Efficient LLM Inference with Low-Bit KV Cache via Outlier-Suppressed Vector Quantization \ The Key-Value (KV) cache introduces substantial memory overhead during largelanguage model (LLM) inference. Although existing vector quantization (VQ)methods reduce KV cache usage and provide flexible representational capacityacross bit-wi… \ Source • arXiv cs.CL • 19:35
• Large Language Models Achieve Gold Medal Performance at the International Olympiad on Astronomy & Astrophysics (IOAA) \ While task-specific demonstrations show early success in applying largelanguage models (LLMs) to automate some astronomical research tasks, they onlyprovide incomplete views of all necessary capabilities in solving astronomyproblems, calli… \ Source • arXiv cs.CL • 17:34
• Paying Attention to Hybrid Attention: Untangling the Issues with Conversion Methods \ Transformers' quadratic computational complexity limits their scalabilitydespite remarkable performance. While linear attention reduces this to linearcomplexity, pre-training such models from scratch remains, in most cases,prohibitively ex… \ Source • arXiv cs.CL • 15:11
• Towards Reliable and Practical LLM Security Evaluations via Bayesian Modelling \ Before adopting a new large language model (LLM) architecture, it is criticalto understand vulnerabilities accurately. Existing evaluations can be difficultto trust, often drawing conclusions from LLMs that are not meaningfullycomparable, … \ Source • arXiv cs.CL • 11:22
• lm-Meter: Unveiling Runtime Inference Latency for On-Device Language Models \ Large Language Models (LLMs) are increasingly integrated into everydayapplications, but their prevalent cloud-based deployment raises growingconcerns around data privacy and long-term sustainability. Running LLMs locallyon mobile and edge … \ Source • arXiv cs.LG • 19:05
• Hybrid Quantum-Classical Policy Gradient for Adaptive Control of Cyber-Physical Systems: A Comparative Study of VQC vs. MLP \ The comparative evaluation between classical and quantum reinforcementlearning (QRL) paradigms was conducted to investigate their convergencebehavior, robustness under observational noise, and computational efficiency ina benchmark control… \ Source • arXiv cs.LG • 17:09
• Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks \ In this paper, we study the surprising impact that truncating text embeddingshas on downstream performance. We consistently observe across 6state-of-the-art text encoders and 26 downstream tasks, that randomly removingup to 50% of embeddin… \ Source • arXiv cs.LG • 15:43
• Conformal Prediction in Hierarchical Classification with Constrained Representation Complexity \ Conformal prediction has emerged as a widely used framework for constructingvalid prediction sets in classification and regression tasks. In this work, weextend the split conformal prediction framework to hierarchical classification,where … \ Source • arXiv stat.ML • 10:10
• CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credits \ Diffusion large language models (dLLMs) generate text through iterativedenoising steps, achieving parallel decoding by denoising only high-confidencepositions at each step. However, existing approaches often repetitively remasktokens due t… \ Source • arXiv cs.CL • 19:08
• Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models \ Reasoning language models improve performance on complex tasks by generatinglong chains of thought (CoTs), but this process can also increase harmfuloutputs in adversarial settings. In this work, we ask whether the long CoTs canbe leverage… \ Source • arXiv cs.CL • 18:30

Big Tech

No items today.

Regulation & Standards

No items today.

Enterprise Practice

No items today.

Open-Source Tooling

No items today.

— Personal views, not IBM. No tracking. Curated automatically; links under 24h old.