Phronix: The Capability-Centric Learning Paradigm for Agents

arXiv cs.AI Papers
agentic-learning llm-agents capability-store scientific-computing mcp-tools self-improvement

Summary

The paper introduces Phronix, a capability-centric learning paradigm that lets LLM-based agent harnesses acquire, validate, and reuse capabilities from their own executions via a predictor-validator-corrector loop, achieving 93-95% first-attempt success on recurring tasks and boosting cheaper models from ~28% to ~90% on SWE-bench.

arXiv:2610.31789v1 Announce Type: new Abstract: Scientific computing increasingly relies on large language model-based agents for performance engineering tasks such as instrumentation, analysis, and workflow porting. However, agent effectiveness depends heavily on the surrounding harness, which typically relies on manually written skills and curated memory. This limits autonomous learning, forcing agents to repeatedly rediscover solutions rather than retain and reuse knowledge from prior executions. We introduce Phronix, a capability-centric learning paradigm that enables agentic harnesses to acquire, validate, and refine reusable capabilities from their own executions. Phronix operates through a predictor-validator-corrector loop: MCP tools and associated skills propose solutions, tool executions validate them, and observed outcomes refine the stored capabilities. It distills every observed interaction into a capability store and researches a failing task rather than abandoning it. First, such learned capabilities solve a recurring task on the first attempt in 93-95% of cases across every model, significantly reducing solution non-determinism and improving reliability. Second, Phronix turns an agent's own execution into 7.6-30.8 durable capabilities per task without user instruction, and creation falls below two once the needed capability exists. Third, warm-start runs consume only 0.38-0.40 million tokens per task, which is three to four times less than a single baseline attempt, while a cold-start costs no more than the baseline. On the 1000 Genome Pegasus workflow, instrumenting a production application for I/O performance analysis, Phronix passes all 11 correctness checks against six for the baseline harness, and also lifts the task solving rates of two cheaper models from 29% and 27% of SWE-bench tasks solved to 91% and 87%, respectively, matching the performance of the most expensive evaluated model.
Original Article

Similar Articles

@xiangxiang103: This new paper from Tsinghua × Microsoft Research Asia — my biggest takeaway after reading it: DSH is seriously underrated! The paper is about HarnessMM, a method that stops agents from "rethinking everything from scratch after each step." It uses Codex and DeepSeek Harness, two industrial-grade harnesses, as experimental subjects…

X AI KOLs Timeline

The HarnessMM paper, a collaboration between Tsinghua University and Microsoft Research Asia, proposes improving agent harness token efficiency through protected decision reuse and recursive self-improvement. Experiments show DeepSeek Harness (DSH) benefits significantly: HumanEval success rate rises from 71.43% to 85.71% with 87.93% fewer tokens per problem, while Codex only saves 29.54%.

Agentic Systems for Calibrating Cosmological Simulations

arXiv cs.CL

The paper proposes AgenticTuner, an agentic pipeline using an LLM as the optimizer to calibrate cosmological simulations by leveraging human-readable priors like codebases and documentation, achieving competitive calibration performance at a fraction of the CPU cost of Bayesian optimization.

Flide: Accelerating Decision Inference with Exact State Forking

arXiv cs.CL

Flide is an efficient and exact inference engine for LLM decision models that decouples logical tree structures from a single flattened token stream via ForkAttention, answering queries up to 9.5x faster than vLLM and SGLang while staying closest to a float32 reference.

KEEPER: Learning Control-Sufficient Textual Memory for Long-Horizon LLM Agents

arXiv cs.CL

The paper introduces KEEPER, a comparative study of four auxiliary objectives for building control-sufficient textual memory in long-horizon LLM agents, finding that reward-related prediction yields the best task performance and introducing self-consistency bias as a measure of error persistence in recursive memory updates.