Rohan Surana | Xunyi's Academic Website

Reasoning Pathologies in Large Language Models: A Diagnostic Perspective

Tue, 01 Sep 2026 00:00:00 +0000

A diagnostic study of reasoning failures in large language models.

The Geometry of Agent Skills: Non-Commutative Composition in Representation Space

Tue, 01 Sep 2026 00:00:00 +0000

A study of the geometry of agent skills and how their composition depends on order in representation space.

Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering

Mon, 20 Jul 2026 00:00:00 +0000

SOPHIA detects when a model repeats unproductive reasoning states and intervenes through hidden-state steering. The method uses transitions between reasoning states to guide generation, improving answer quality and token efficiency.

RRPO: Reference-Relative Policy Optimization with Stratified Conditional Rollouts

Mon, 20 Jul 2026 00:00:00 +0000

RRPO extends group-relative reinforcement learning to tasks without a direct correctness verifier. It compares candidate rollouts against positive and negative reference sets to construct training advantages for reasoning and open-ended generation.

Skill-R1: Agent Skill Evolution via Reinforcement Learning

Sun, 10 May 2026 00:00:00 +0000

Skill-R1 trains a small generator to improve reusable natural-language procedures while keeping the task model fixed. Feedback from successive rollouts guides skill revisions, enabling adaptation with both open and closed models.

Generate, Filter, Control, Replay: A Comprehensive Survey of Rollout Strategies for LLM Reinforcement Learning

Wed, 08 Apr 2026 00:00:00 +0000

This survey organizes rollout design for language-model reinforcement learning into generation, filtering, control, and replay. It connects design choices to reliability, coverage, and computational cost, with examples across reasoning and agent tasks.