#machine-learning
Every summary, chronological. Filter by category, tag, or source from the rail.
ε-MemEvo: Adaptive Cross-Task Memory Transfer for LLM Evolution
ε-MemEvo improves LLM-based program evolution by using an adaptive memory transfer mechanism that selectively reuses successful code patterns across different tasks, significantly increasing search efficiency.
Trie Automata for Efficient Constrained Decoding
Trie automata provide a memory-efficient and performant method for enforcing complex constraints during LLM decoding, particularly when dealing with massive sets of valid output tokens.
Measuring and Restoring Constraint Influence in LLMs
LLMs often ignore complex constraints in long dialogues, treating them as 'dead text.' This research introduces a method to quantify and restore constraint adherence in black-box models.
CAS: A Causal Attribution Score for Explainable AI
The Causal Attribution Score (CAS) provides a unified framework for evaluating AI model interpretability by measuring the causal impact of features on predictions, bridging the gap between local and global explanations.
LLMs Hit a Hard Limit on Multi-Constraint Instruction Following
LLMs exhibit 'phase transitions' in performance, where adding a single additional constraint causes a sudden, catastrophic drop in instruction-following capability rather than a gradual decline.
Meta-LoRA: Efficient Cross-Domain LLM Personalization
Meta-LoRA enables LLMs to adapt to user preferences across different domains by learning a meta-adapter that generalizes personalization patterns, reducing the need for domain-specific fine-tuning.
Governed Persistent Memory for Long-Horizon AI Agents
This research introduces a 'Governed Persistent Memory' framework that uses source-bound state semantics and fail-closed release mechanisms to improve reliability and safety in long-horizon AI agents.
MindMemOS: A Self-Evolving Memory Layer for AI Agents
MindMemOS introduces a portable, self-evolving memory operating layer that decouples agent intelligence from long-term storage, enabling persistent, adaptive memory across diverse AI architectures.
Dual-Flow Transformers: Decoupling Prefill and Decode Paths
Dual-Flow Transformers optimize LLM inference by decoupling the primary prefill path from additional decode-time computation, allowing for more efficient resource allocation during the two distinct phases of generation.
Aligning AI with Human Reasoning Processes
Current AI alignment methods focus on outcomes rather than cognitive processes. To build reliable systems, we must shift toward alignment techniques that mirror human reasoning, ensuring models arrive at conclusions through transparent, human-compatible logic.
Language-Dependent Safety: How Non-English Prompts Alter LLM Behavior
Research indicates that LLMs exhibit varying safety alignment levels across languages, with non-English prompts—specifically Japanese—often triggering more cautious responses to harmful queries compared to English.
Modular Prompt Optimization: Improving LLM Performance via Segmentation
Moving from monolithic prompt optimization to segment-level modularity allows for more precise, interpretable, and effective tuning of LLM instructions.
Forecasting Side Effects of Activation Steering
Activation steering allows for precise control over LLM behavior, but it often introduces unintended side effects. This research provides a framework to predict these downstream behavioral changes before deployment.
Automating Process Engineering Diagrams with LLMs
This research explores a multi-agent framework for generating and validating Process Flow Diagrams (PFDs) and Piping and Instrumentation Diagrams (P&IDs) using LLMs to reduce manual engineering errors.
Mitigating Bus Bunching via Reinforcement Learning and Semantic Embeddings
This research introduces a reinforcement learning framework that uses semantic stop embeddings to predict and prevent bus bunching, significantly improving transit reliability compared to traditional control methods.
Training Data Granularity and Parametric Modularity in LLMs
The research establishes that the granularity of training data directly dictates whether knowledge within an LLM is modular and detachable, or merely decodable but entangled.
Sparse Coding for Latent Communication in VLM Agents
This paper introduces a post-hoc sparse coding method to interpret and analyze the latent communication signals exchanged between vision-language model (VLM) agents, providing a framework for understanding multi-agent internal states.
Evaluation-Conditioned Training for Stronger Oversight
Evaluation-Conditioned Training (ECT) improves model performance by training agents to adapt their behavior based on the strength of the oversight regime they operate under, ensuring better generalization.
Automating LLM Adversarial Attacks with GFlowNets
Generative Flow Networks (GFlowNets) provide a more efficient, diverse, and scalable framework for discovering adversarial prompts compared to traditional gradient-based or evolutionary search methods.
SBCO: Self-Supervised Verifier-Grounded Harness Optimization
SBCO is a framework for optimizing planning agents by using self-supervised, verifier-grounded harness optimization to improve decision-making accuracy without requiring extensive human-labeled data.
MESA: Task-Adaptive Evidence Selection for Agent Memory
MESA improves long-horizon agent performance by using a task-adaptive, multi-structure memory selection framework that retrieves relevant evidence more effectively than standard retrieval methods.
CHORUS: Improving Testbench Coverage via Complementary AI Experts
CHORUS improves hardware verification by using a multi-expert AI framework to generate diverse, high-coverage testbench stimuli, outperforming single-model approaches.
MIDAS: Handling Incomplete Multimodal Sentiment Analysis
The MIDAS framework addresses incomplete multimodal data by disentangling shared and private information while using uncertainty-aware fusion to maintain sentiment prediction accuracy when modalities are missing.
Automating Behavioral Research for AI Agents
This paper introduces a framework for scaling behavioral scientific research on AI agents, moving beyond manual evaluation to automated, reproducible experimental pipelines.
Quantifying the Carbon Footprint of Deep Learning Models
This review analyzes the environmental impact of deep learning, highlighting the massive carbon costs of training large models and proposing strategies for more sustainable AI development.
Continuously Improving AI Agents via Trace Data Mining
To improve autonomous agents, treat them like machine learning models: collect execution traces, mine them for feedback, and use that data to iteratively refine prompts, fine-tune models, and update agent state.
AI EngineerContinual Learning via Distillation: A 2x2 Taxonomy
Enterprises can implement continual learning by mapping distillation tasks across a 2x2 grid of offline/online traces and hints, allowing for immediate performance improvements without requiring 'golden' datasets.
Democratizing Frontier AI: Automating Discovery and Scaling
The era of massive, monolithic pre-training is hitting a ceiling. By automating model training and data optimization, we can shift the focus from compute-heavy scaling to domain-specific innovation, allowing more builders to participate at the frontier.
Scaling Expertise: Moving Beyond Raw Intelligence in AI Agents
Current AI agents excel at symbolic tasks like coding but struggle with real-world digital work because they lack 'expertise'—the ability to learn and adapt to idiosyncratic micro-worlds through continuous learning.
Scaling Compute on Context: Moving Beyond Public Data
Current AI models excel on public data but fail to acquire deep, personalized knowledge. The solution lies in 'scaling compute on context'—using recursive self-improvement to deepen a model's understanding of private data without hitting a synthetic data wall.
Showing 30 of 472