AI Post Transformers

This October 5 2025 paper presents the first mechanistic explanation for a persistent training instability experienced when using low-precision arithmetic (specifically BF16) with the Flash Attention algorithm in transformer models. The paper identifies the core problem as a "catastrophic loss explosion" caused by two interacting phenomena: the emergence of similar low-rank representations within the attention mechanism and the accumulation of biased rounding errors inherent to BF16 addition during the attention output calculation. This bias leads to a systematic error in the gradient updates, causing the spectral norm of weights to increase and derailing the training process. To validate this analysis, the authors introduce a minimal modification to the softmax computation in Flash Attention that mitigates the rounding bias and successfully stabilizes the training, offering a practical solution to this long-standing issue. Source: https://arxiv.org/pdf/2510.04212

On October 6, 2925 Anthropic introduces Petri (Parallel Exploration Tool for Risky Interactions), an open-source framework developed for automated auditing to accelerate AI safety research. Petri uses AI-driven auditor agents to interact with and test the behavior of target language models across diverse, multi-turn scenarios, automating the process of environment simulation and initial transcript analysis. A judge component then scores the generated transcripts across dozens of dimensions, such as "unprompted deception" or "whistleblowing," to quickly surface misaligned behaviors like autonomous deception and cooperation with misuse. The text provides a detailed technical overview of Petri's architecture, including how researchers form hypotheses, create seed instructions, and utilize the automated assessment and iteration steps, while also discussing the limitations and biases found in the auditor and judge agents during pilot evaluations. Source: https://alignment.anthropic.com/2025/petri/

This October 6, 2025 paper from Alexia Jolicoeur-Martineau at Samsung SAIL Montréal, provides an overview and detailed comparison of two recurrent reasoning models: the Hierarchical Reasoning Model (HRM) and the proposed Tiny Recursive Model (TRM). HRM is a complex, biologically-inspired approach that uses two small neural networks and deep supervision to outperform Large Language Models (LLMs) on difficult puzzle tasks like Sudoku and ARC-AGI. The authors introduce TRM as a simpler, more efficient alternative that utilizes a single tiny, two-layer network to achieve superior generalization while using significantly fewer parameters than both HRM and powerful LLMs. The document highlights that TRM simplifies HRM by removing the need for fixed-point theorems and complex biological justifications, instead relying on recursive reasoning to progressively refine its predicted answer, leading to state-of-the-art results on several benchmarks.Sources:https://arxiv.org/html/2510.04871v1https://www.artificialintelligence-news.com/news/samsung-tiny-ai-model-beats-giant-reasoning-llms/

The October 2025 papar provide an overview of Agent Context Optimization (ACON), a novel framework designed to enhance the efficiency and performance of Large Language Model (LLM) agents operating in complex, long-horizon tasks. ACON addresses the challenge of unbounded context growth—which increases costs and reduces effectiveness—by optimally compressing both environment observations and interaction histories into concise summaries. The framework uses a gradient-free guideline optimization pipeline where a capable LLM analyzes compression failures from contrastive trajectories to refine the compression instructions in natural language. Furthermore, the optimized compressor can be distilled into smaller models to reduce computational overhead, with empirical results demonstrating significant reductions in peak tokens and memory usage while preserving or even improving task accuracy across multiple benchmarks. Source: https://arxiv.org/pdf/2510.00615

Thus November 2024 paper and new analysis in September 2025 provide a comprehensive overview of a novel Analog In-Memory Computing (AIMC) architecture designed to accelerate the attention mechanism in Large Language Models (LLMs). The core technology involves using capacitor-based gain cells (made from emerging OSFETs like IGZO) to store the Key (K) and Value (V) projections of the KV cache directly within the memory arrays, enabling parallel, analog dot-product computation that drastically reduces the latency and energy consumed by data movement in traditional GPUs. Simulations indicate performance improvements of up to 7,000× speedup and 90,000× energy reduction compared to NVIDIA A100 GPUs for the attention step alone, and the research introduces a hardware-aware training methodology to maintain accuracy despite analog non-idealities and the use of a simplified ReLU-based activation function instead of softmax. The text also notes that while major chipmakers are engaged in tangential AIMC research, this specific attention mechanism design is currently a prototype from academic institutions and faces a multi-year timeline for commercial readiness and scaling to trillion-parameter models.Sources:https://arxiv.org/pdf/2409.19315https://www.nextbigfuture.com/2025/09/analog-in-memory-computing-attention-mechanism-for-fast-and-energy-efficient-large-language-models.html

The October 2025 paper introduces CoDA (Collaborative Data-visualization Agents), a novel multi-agent system designed to automate complex data visualization from natural language queries, addressing the limitations of existing rule-based and single Large Language Model (LLM) approaches. The core innovation of CoDA is its collaborative paradigm, where specialized LLM agents—focused on tasks like query analysis, data processing, design mapping, and self-reflection—work together through an iterative refinement loop to enhance output quality and robustness. Experimental results demonstrate that CoDA significantly outperforms state-of-the-art baselines (MatplotAgent, VisPath, CoML4VIS) on benchmarks like MatplotBench and Qwen Code Interpreter, achieving superior execution pass rates and visualization success rates, particularly when dealing with complex queries, multi-file data, and specific stylistic constraints. Ablation studies further validate the necessity of CoDA’s architectural components, such as the Global TODO List and the Search Agent, confirming that structured planning and external knowledge retrieval are crucial for overcoming ambiguity and ensuring high-fidelity code generation. The paper concludes that this agentic approach transforms visualization generation into a more resilient and adaptive problem-solving process, making it effective for real-world data science tasks. Source: https://arxiv.org/pdf/2510.03194

We compare and contrast the math behind two recent research papers which we have covered individually before on this podcast: July 2025: Learning without training: The implicit dynamics of in-context learning https://arxiv.org/pdf/2507.16003 September 2025: Federated Learning with Ad-hoc Adapter Insertions: The Case of Soft-Embeddings for Training Classifier-as-Retriever https://arxiv.org/pdf/2509.16508 The first source explores the concept of In-Context Learning (ICL) in neural networks, proposing that the effect of context on a token's output is equivalent to an implicit weight update in the neural network, specifically in the MLP layer, generalizing the transformer block using a contextual block notion. This work provides an explicit low-rank update formula for this implicit weight modification and mathematically demonstrates that token consumption aligns with an implicit gradient descent learning dynamics on the network weights. The second source introduces a novel retrieval-augmented generation (RAG) architecture called Classifier-as-Retriever (CaR) for memory-constrained edge devices, proposing to use a frozen Small Language Model (SLM) augmented with a small trainable adapter network to generate "soft embeddings" and a trainable classifier head instead of conventional similarity functions. Crucially, this architecture is designed for distributed training using Federated Learning (FL), incorporating Differential Privacy (DP) techniques to ensure client-side data protection and demonstrating significant speedup advantages over centralized training.

The September 29 2025 paper introduces DC-VideoGen, a new post-training framework designed to significantly accelerate video diffusion models and reduce their training costs. This system relies on two main innovations: the Deep Compression Video Autoencoder (DC-AE-V), which achieves high spatial and temporal compression using a novel chunk-causal temporal modeling approach to maintain reconstruction quality; and AE-Adapt-V, an efficient finetuning strategy using LoRA to adapt pre-trained models to the new latent space while preserving their original knowledge and semantics. Experimental results demonstrate that DC-VideoGen successfully accelerates inference speed by up to 14.8× for high-resolution videos and drastically reduces training expenses, all while maintaining or improving video generation quality across tasks like text-to-video and image-to-video generation. Source: https://arxiv.org/pdf/2509.25182

The sources (October 2022, March 2025) provide an extensive examination of emergent abilities in large language models (LLMs), defining them as unpredictable, sharp performance increases on specific tasks that occur only after models reach a critical scale. The initial source establishes this concept through empirical evidence on benchmarks like BIG-Bench, showing tasks where performance jumps suddenly from near-random, particularly in few-shot prompting and specialized prompting techniques like Chain-of-Thought. The subsequent survey source expands on this by framing emergence within the broader context of in-context learning, discussing how factors like model quantization, task complexity, and pre-training loss thresholds influence the appearance of these abilities. Both sources acknowledge the ongoing debate about whether these sudden leaps are genuine phenomena or merely artifacts of evaluation metrics that do not award partial credit, while also highlighting the emergence of harmful behaviors and advanced reasoning capabilities in LLM-powered AI agents as scale increases.Sources:https://arxiv.org/pdf/2206.07682https://arxiv.org/pdf/2503.05788

The November 2024 paper introduces GNN101, an open-source, web-based interactive visualization tool designed to help non-experts learn about Graph Neural Networks (GNNs), whose complex nature often challenges beginners. This educational tool addresses limitations in existing resources by seamlessly integrating mathematical formulas with visualizations across multiple abstraction levels, from a model overview to detailed matrix calculations. GNN101 features complementary views—a node-link diagram for intuitive graph understanding and a matrix view for a comprehensive feature overview—to illustrate how GNNs process graph data and update node features. The authors detail the design goals, implementation, and initial deployment of GNN101, showing its usability and effectiveness in making GNN computations more intuitive and engaging for students. Source: https://arxiv.org/html/2411.17849v1