FastCode: Fast and Cost-Efficient Code Understanding and Reasoning

Zhonghang Li; Zongwei Li; Yuxuan Chen; Han Shi; Jiawei Li; Jierun Chen; Haoli Bai; Chao Huang

cs.SEMar 1, 2026

FastCode: Fast and Cost-Efficient Code Understanding and Reasoning

Authors: Zhonghang Li, Zongwei Li, Yuxuan Chen, Han Shi, Jiawei Li, Jierun Chen, Haoli Bai, Chao Huang

Organizations: The University of Hong Kong · Huawei Noah’s Ark Lab

Abstract

Repository-scale code reasoning is a cornerstone of modern AI-assisted software engineering, enabling Large Language Models (LLMs) to handle complex workflows from program comprehension to complex debugging. However, balancing accuracy with context cost remains a significant bottleneck, as existing agentic approaches often waste computational resources through inefficient, iterative full-text exploration. To address this, we introduce FastCode, a framework that decouples repository exploration from content consumption. FastCode utilizes a structural scouting mechanism to navigate a lightweight semantic-structural map of the codebase, allowing the system to trace dependencies and pinpoint relevant targets without the overhead of full-text ingestion. By leveraging structure-aware navigation tools regulated by a cost-aware policy, the framework constructs high-value contexts in a single, optimized step. Extensive evaluations on the SWE-QA, LongCodeQA, LOC-BENCH, and GitTaskBench benchmarks demonstrate that FastCode consistently outperforms state-of-the-art baselines in reasoning accuracy while significantly reducing token consumption, validating the efficiency of scouting-first strategies for large-scale code reasoning. Source code is available at https://github.com/HKUDS/FastCode.

Figures & tables

Panel of Figure 1 — Figure 1: Accuracy and cost trade-off comparison

Figure 2: The overall framework of FastCode.

Metric	Community Baselines				Commercial Tools					Ours
Metric	Direct LLM	Func RAG	Sliding RAG	SWEQA-Agent	DeepWiki	CodeWiki	Gemini Code	Claude Code	Cursor	FastCode
Correctness	7.36	7.96	8.06	8.36	8.02	7.89	8.03	8.31	8.30	8.44
Completeness	3.71	5.75	5.99	6.94	5.90	6.49	6.09	7.25	8.13	7.94
Clarity	7.38	8.97	9.01	9.10	8.73	8.75	8.97	8.76	8.71	8.84
Relevance	8.94	9.56	9.65	9.72	9.36	9.21	9.68	9.56	9.39	9.35
Reasoning	4.91	7.38	7.51	8.22	7.52	7.82	7.92	8.21	8.64	8.72

Table 1: Comparison of baseline performance on the SWE-QA dataset (3-repository subset).

LongCodeQA	32K	64K	128K	256K	512K	1M
Jamba 1.5 - 400B Large	69.0	69.7	72.8	54.2	-	-
Llama 3.1 - 405B Ins.	69.9	72.4	67.4	-	-	-
Llama 4 Scout	66.4	73.7	70.7	63.1	78.7	76.0
GPT-4o	65.5	76.3	74.3	-	-	-
GPT-4.1	72.6	73.7	78.3	72.3	78.7	80.0
Qwen2.5-14B Ins.	61.9	65.8	68.5	63.1	70.2	40.0

Table 2: Performance comparison on LongCodeQA benchmark.

Figure 3: File-level localization accuracy (Acc@K) on Loc-Bench.

Figure 4: Performance of different frameworks on GitTaskBench.

Panel of Figure 5 — Figure 5: Performance-cost tradeoffs for baselines and our framework across four benchmarks.

Figure 6: Ablation results on SWE-QA.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Provider	Price ($ / 1M)
Model	Provider	Input	Output
Gemini-3-Flash	Poe	0.40	2.40
Gemini-2.5-Pro	Poe	0.87	7.00
Claude 3.5 Sonnet	Poe	2.60	13.00
Claude 3.7 Sonnet	Poe	2.60	13.00
Qwen3-Coder-30B	OpenRouter	0.07	0.27

Appendix

Table 3: Detailed API usage costs. Prices are denoted in USD per 1M tokens.

Method	Total Score	Cost
Claude Code	42.08	0.057
Cursor	43.17	0.071
FastCode	43.28	0.032
Improvement over Claude Code	+2.85%	-43.86%

Appendix

Table 4: SWE-QA Performance Comparison

Context Length	Accuracy (%)		Cost		Acc. Gain
Context Length	Direct	+FastCode	Direct	+FastCode	Acc. Gain
32K	65.5	69.9	0.083	0.041	+6.72%
64K	69.7	69.7	0.166	0.045	+0.00%
128K	71.7	72.83	0.333	0.069	+1.58%
256K	66.6	78.5	0.666	0.057	+17.87%
512K	-	76.6	1.331	0.075	-

Appendix

Table 5: LongCodeQA Performance: Direct vs. FastCode-Enhanced

Method	Acc@1 (%)	Acc@3 (%)	Acc@5 (%)	Cost
Existing Tools (Claude 3.5 Sonnet)
MoatlessTools	72.63	85.77	86.13	0.46
SWE-agent	77.37	87.23	90.15	0.67
Openhands	76.28	89.78	90.15	0.79
LocAgent	77.74	91.97	94.16	0.66
LocAgent (Fine-tuned Qwen)

Appendix

Table 6: LOC-BENCH (SWE-Bench-Lite) Performance

Method	ECR (%)	TPR (%)	Cost	Cost Ratio
Aider (Claude 3.5)	16.67	12.96	0.0304	1.00 $\times$
SWE-Agent (Claude 3.5)	41.67	22.23	1.38	45.39 $\times$
SWE-Agent (Claude 3.7)	64.81	42.59	1.67	54.93 $\times$
OpenHands (Claude 3.5)	53.70	40.74	8.95	294.41 $\times$
OpenHands (Claude 3.7)	72.22	48.15	29.8	980.26 $\times$
FastCode (Claude 3.5)	61.11	46.30	0.0878	2.89 $\times$

Appendix

Table 7: GitTaskBench Performance Comparison

Base Model	Direct LLM	SWEQA-Agent	FastCode
Ministral-3b	27.39	33.45	36.13
Deepseek-r1-8b	30.11	34.46	38.56
Qwen3-Coder-30b-A3b	35.77	37.08	41.44
Gemini-3-Flash	32.65	42.47	43.44

Appendix

Table 8: FastCode performance on SWE-QA under different LLMs.

Repository	DeepSeek-R1-8B	Gemini-3-Flash	Ministral-3B	Qwen3-Coder-30B
astropy	27.58	32.42	25.60	34.08
sphinx	28.54	32.67	24.60	34.44
matplotlib	31.52	33.88	28.38	37.48
pylint	28.46	32.79	27.33	35.56
flask	33.46	32.12	27.81	36.42
pytest	28.54	32.50	26.75	35.46

Appendix

Table 9: Direct LLM results on SWE-QA across all repositories with different backbone LLMs.

Repository	DeepSeek-R1-8B	Gemini-3-Flash	Ministral-3B	Qwen3-Coder-30B
astropy	31.44	41.32	35.15	36.21
sphinx	32.71	42.08	29.35	38.40
matplotlib	35.12	43.48	33.75	37.67
pylint	34.29	42.60	34.85	37.12
flask	35.71	42.15	38.67	37.60
pytest	33.25	42.92	27.44	39.77

Appendix

Table 10: SWEQA-Agent results on SWE-QA across all repositories with different backbone LLMs.

Repository	DeepSeek-R1-8B	Gemini-3-Flash	Ministral-3-3B	Qwen3-Coder-30B
astropy	36.40	41.58	33.88	40.33
sphinx	37.81	44.10	35.96	41.75
matplotlib	40.56	43.44	38.73	41.79
pylint	37.98	43.42	34.06	40.65
flask	40.27	44.98	38.12	43.19
pytest	38.65	44.12	35.00	41.96

Appendix

Table 11: FastCode results on SWE-QA across all repositories with different backbone LLMs.

Efficient Language Model Reasoning

Efficient language model reasoning studies how to improve problem-solving accuracy while reducing inference, memory, latency, and training costs.

See all 370 papers

CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents
Mar 18, 2026Lintang Sutawika, Aditya Bharat Soni, Bharath Sriraam R R +8Code Localization Code Search
Carnegie Mellon University · Independent · OpenHands
DeepRepoQA: Code Repository Question Answering with Deep Agent Exploration
Aug 25, 2026Weihan Peng, Yuling Shi, Yingwei Ma +3Question Answering Software Engineering Agents
Shanghai Jiao Tong University · The Hong Kong University of Science and Technology · University of California San Diego
Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories
Oct 21, 2026Sujoy Roychowdhury, Joseph H. R. Isaac, Jiangning Gao +2Cross-Modal Retrieval Semantic Search
Ericsson R&D

Mar 18, 2026J/K move · Enter open · S save

CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents

Lintang Sutawika, Aditya Bharat Soni, Bharath Sriraam R R, Apurva Gandhi+7

Carnegie Mellon University · Independent · OpenHands

Code Localization Code Search RL for LLM Agents Reinforcement Learning Information Retrieval

Open paper AskPDF arXiv

A prerequisite for coding agents to perform tasks on large repositories is code localization - the identification of relevant files, classes, and functions to work on. While repository-level code localization has been performed using embedding-based retrieval approaches such as vector search, recent work has focused on developing agents to localize relevant code either as a standalone precursor to or interleaved with performing actual work. Most prior methods on agentic code search equip the agent with complex, specialized tools, such as repository graphs derived from static analysis. In this paper, we demonstrate that, with an effective reinforcement learning recipe, a coding agent equipped with nothing more than a standard Unix terminal can be trained to achieve strong results. Our experiments on three benchmarks (SWE-Bench Verified, Pro, and Lite) reveal that our models consistently achieve superior or competitive performance over 2-18x larger base and post-trained LLMs and sometimes approach performance provided by closed models like Claude Sonnet, even when using specialized scaffolds. Our work particularly focuses on techniques for re-purposing existing coding agent environments for code search, reward design, and RL optimization. We release the resulting model family, CodeScout, along with all our code and data for the community to build upon.

Abstract

Efficient Language Model Reasoning

Explore similar work

CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents

DeepRepoQA: Code Repository Question Answering with Deep Agent Exploration

Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories