The Morning
From the arXiv
Reconciling Process Supervision with Outcome-Based Credit in Agentic Policy Optimization
his paper introduces TASPO, a method to bridge the gap between outcome-based credit and process supervision in reinforcement learning for language agents. TASPO converts privileged training information into actionable credit for individual decisions, ensuring that each action is appropriately rewarded based on its contribution to the overall task outcome, rather than uniformly distributing credit across entire trajectories.
Scaling Large Reasoning Models beyond Human Supervision: A Path toward Superintelligence
This paper proposes a method for scaling large reasoning models (LRMs) beyond human supervision by gradually reducing reliance on direct human feedback. The core idea is to transition from human judgments to automated verifiers and self-generated learning curr…
Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning
This paper proposes "knowledge-aligned SFT" to reduce hallucinations in language models by ensuring fine-tuning targets are consistent with the model's existing knowledge. The authors introduce new methods like "Recall Rewrite" which only keeps claims the mode…
A Model with No Head and Many Thoughts
This paper introduces Soft Latent Thinking, a method that replaces the computationally expensive LM head during reasoning with a lightweight projector. This allows reasoning to occur in continuous embedding space, avoiding discrete tokenization at each step. T…
Low-Resource Preference Adaptation of LLMs via Activation-Based Label Propagation
This paper proposes a method to adapt Large Language Models (LLMs) to user preferences with minimal labeled data. It leverages the observation that LLM activations for preferred and rejected responses form distinct clusters, even in pre-trained models. By trai…
S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?
This paper introduces S3Gym, a benchmark designed to evaluate if Large Language Models (LLMs) can improve themselves. S3Gym tests LLMs' abilities in self-testing, self-judging, and…
A Universal Context-Reuse Layer for Cross-Model KV Sharing
This paper introduces a "Universal Context-Reuse Layer" that enables KV-cache sharing between different Large Language Models (LLMs), even those with varying architectures or token…
HSRM: Hidden-State Reward Models for Test-Time Verification
This paper introduces HSRM, a method for verifying mathematical reasoning generated by large language models. Instead of re-processing the text of candidate solutions, HSRM directl…
Learning to Evaluate Before Improving: Automatic Rubric Induction for Automatic Research Agents
This paper introduces AutoSciRub, a novel framework for autonomous research agents. AutoSciRub's core method is to first automatically generate an executable rubric that defines sp…
LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation
This paper introduces LightNav-0, a generalist embodied navigation model that directly leverages the spatial reasoning capabilities of pre-trained Vision-Language Models (VLMs). It…
The Town Square
Apple is experiencing unexpected high demand for its Mac Mini and Mac Studio, driven by AI workloads, catching them off guard and impacting supply.
Workshops
Archify is an agent skill that generates beautiful, verifiable architecture and workflow diagrams as self-contained HTML, offering motion and crisp export capabilities.
This repository provides a comprehensive library of validated AI skills and scientific databases, enabling any AI agent to function as a scientist across various disciplines like biology and chemistry.
