The Morning
From the arXiv
A Verifier-Guided Explainable Reasoning Framework with Gold-Anchored QLoRA, Task-Aware Mixture-of-Experts, and Group-Relative RLVR
his paper introduces a framework to improve the explainability and verifiability of LLM reasoning in educational question answering. It achieves this by using a gold-anchored QLoRA for initial model adaptation, a task-aware router to direct problems to specialized symbolic solvers (like Z3), and a reinforcement learning approach (RLVR) that incorporates feedback from these solvers to refine the LLM's explanations. The core contribution is a more transparent and trustworthy reasoning process for LLMs, particularly in educational contexts.
ACE: Adaptive Calibration-Free Expert Skipping for MoE-based LLMs
ACE addresses redundant computation in MoE LLMs by adaptively skipping experts for each token without requiring extra training or calibration. It uses a Global Spectral Proxy to estimate expert capacity and Router-Conditioned Refinement to evaluate expert resp…
RISE: Recursive Improvement via Self-Extrapolating Policy Distillation
RISE improves language model training by creating a synthetic teacher from the model's own experience, bypassing the limitations of external teachers or in-context learning. It extrapolates parameter or logit changes from the model's training trajectory to gen…
Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents
Trace2Tower addresses limitations in LLM agent skill learning by introducing a transition-aware framework that distills raw execution traces into a hierarchical skill structure. It achieves this by abstracting interactions into canonical events, building a gra…
Unifying ICL, SFT, KL-Regularized RL Through a Bayesian Lens
This paper unifies diverse LLM training and evaluation methods like ICL, SFT, and RL under a Bayesian framework. The core method involves constructing a generalized Bayesian posterior over outputs using a prior model and a utility signal, then approximating th…
A Structured Debate-Mixture-of-Agents Framework for Complex Clinical Diagnostic Decision Support
This paper introduces Debate-Mixture-of-Agents (DMoA), a novel multi-agent framework that mimics structured clinical diagnostic reasoning through role-based interactions. DMoA sign…
Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability
This paper investigates how well an agent's memory survives model upgrades by comparing four memory storage methods: raw history, RAG chunks, natural language notes, and a fixed-sc…
GUT: Quantifying and Optimizing the Reasoning Uncertainty of LLMs via Graph Complexity
This paper introduces GUT, a method to quantify and reduce reasoning uncertainty in LLMs. GUT models LLM reasoning as a directed acyclic graph, capturing all possible branches. It …
Measuring AI Accountability Through Argumentation Analysis: Can Model Reasoning Withstand Scrutiny?
This paper proposes a novel method for measuring AI accountability by analyzing the structural quality of a model's defense for its verdicts, rather than relying on ground truth. U…
PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting
This paper introduces PRICE, a systematic framework for adapting Large Language Models (LLMs) for Bitcoin price forecasting. It combines efficient fine-tuning (LoRA), recursive inf…
The Town Square
This article argues that using LLMs to write content exposes one's intellectual vulnerabilities and biases, akin to an "open fly."
Workshops
This repository provides a curated collection of "skills" or code snippets designed for real-world engineering tasks, directly sourced from the author's personal `.agents` directory.
This repository provides 38 self-contained HTML+SVG editorial diagram types for Claude Code, Codex, and Pi, designed for clean, shadow-free output without Mermaid dependencies.
