#llm
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.
Reasoning Jury: Improving LLM Evaluation via Multi-Model Consensus
The 'Reasoning Jury' framework improves the reliability of evaluating LLM reasoning traces by using a multi-model consensus approach, reducing the bias and inconsistency inherent in single-model evaluation.
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.
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.
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.
Kog Optimizes GPU Inference Through Low-Level Software Engineering
French startup Kog is challenging the notion that GPUs are poorly suited for agentic AI workloads by using low-level assembly and binary-level optimization to unlock massive inference speed gains on existing datacenter hardware.
Why Computer-Use Models Will Agentify the Web
The web was built for human eyes, not APIs. Instead of waiting for a universal API layer, AI agents will 'agentify' the web by interacting directly with pixels and DOMs, treating browsers as game engines to perform tasks.
Industrial AI Scaling, Local Models, and Cybersecurity Risks
The panel discusses the shift toward industrial-scale AI infrastructure, the rise of high-performance local models like Meta's Muse Glimmer, and the emerging cybersecurity implications of autonomous agent capabilities in upcoming models like OpenAI's Astra.
Optimizing Agentic Workflows with GPT-5.6
GPT-5.6 shifts the economics of agentic AI by enabling high-performance results with smaller models, reduced reasoning effort, and new API primitives like programmatic tool calling and multi-agent orchestration.
Scaling Frontier Intelligence: GPT-5.6 Sol at 750 Tokens/Second
OpenAI is introducing 'Ultrafast' mode, a new service tier powered by Cerebras that enables GPT-5.6 Sol to generate up to 750 tokens per second—a 14x speed increase over standard processing—without sacrificing model intelligence.
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.
Optimizing AI Harnesses to Slash Enterprise Token Costs
Writer’s new Palmyra X6 model and upgraded agentic harness aim to reduce enterprise AI costs by up to 50% by focusing on infrastructure efficiency rather than just model selection.
Google 'All Things Agentic' Hackathon Overview
Google is hosting a global hackathon with $180,000 in prizes, challenging developers to build autonomous, production-ready AI agents using Gemini 3.5 and Google Cloud.
TRACE: A Framework for Trustworthy RAG Systems
The TRACE framework addresses reliability in retrieval-augmented generation by implementing a multi-stage verification process to mitigate hallucinations and ensure factual grounding in conversational AI.
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.
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 EngineerThe Evolution and Future of AI Memory Systems
AI memory has evolved from simple thread-based context to asynchronous 'running profiles.' Despite this, current systems suffer from staleness, lack of cross-platform integration, and an inability to reason over external data sources like email or calendars.
Continual 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.
Building an Automated LLM-Powered Knowledge Base
Transform disorganized raw notes into a structured, interconnected wiki using voice dictation, LLM-based enrichment, and automated cloud-based pipelines.
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.
Scaling Continual Learning with On-Policy Self-Distillation
On-Policy Self-Distillation (OPSD) enables models to learn continuously from real-world data by using privileged hints to guide training, overcoming the infrastructure and reward-density limitations of traditional RLHF and GRPO.
Detecting LLM Hallucinations via Internal State Probing
LLMs often express high confidence in incorrect answers, but internal state probes can detect these errors before the model generates the output, revealing a 'knowing-saying gap'.
NL2SHACL-Bench: Evaluating LLM Performance on SHACL Generation
NL2SHACL-Bench provides a standardized benchmark suite to evaluate how effectively Large Language Models can translate natural language requirements into SHACL (Shapes Constraint Language) for RDF data validation.
Moving Beyond Prompt Engineering: The Power of Context Engineering
Context engineering is the practice of curating and structuring the information environment provided to an LLM, moving beyond simple prompt phrasing to improve reasoning and reduce 'context rot'.
IBM TechnologyShowing 30 of 1201