AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery

Jie Huang; Yuecheng Hong; Yao Chen; Hao Zhong; Linfeng Xiao; Xudong Liu; Qiang Ye; Hao Xia

cs.AIOct 10, 2026

AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery

Authors: Jie Huang, Yuecheng Hong, Yao Chen, Hao Zhong, Linfeng Xiao, Xudong Liu, Qiang Ye, Hao Xia,

Organizations: Harbin Institute of Technology · Independent Researcher · University of Science and Technology of China

Abstract

Formulaic alpha mining seeks to discover executable expressions that capture predictive market signals in a vast, noisy search space. Recent LLM-based approaches use prompting, iterative search, or seed-factor evolution, leaving direct factor construction through policy learning less explored. To address this, we introduce AlphaRLM, a reinforcement-learning framework that trains LLMs to construct executable symbolic factors from financial literature without input factor expressions. AlphaRLM incorporates Literature-Guided Operator Assembly (LEGO), which conditions factor construction on literature-derived ideas and executable operators, providing semantic priors for exploration. It further aligns the generation policy with market feedback through executable-validity and grouped-return-separation rewards that capture non-monotonic structures across factor quantiles. Extensive experiments on CSI 300, CSI 1000, and S&P 500 show that AlphaRLM outperforms strong factor-mining and LLM-based baselines in annualized excess return. Notably, the 8B model achieves strong cross-market performance, highlighting the promise of literature-conditioned, market-aligned factor generation.

Figures & tables

Figure 1: Overview of AlphaRLM. (A) Literature-Guided Operator Assembly (LEGO) combines paper titles and abstracts with permitted market variables and operators to guide factor construction. (B) Executable-validity and grouped-return separation rewards guide DAPO updates to the generation policy. (C) Generated factors undergo screening and training-set XGBoost feature-importance selection, followed by downstream return prediction and portfolio backtesting.

Factor Libraries
	CSI 300					CSI 1000					S&P 500
Method	AR	IC	RankIC	ICIR	MDD	AR	IC	RankIC	ICIR	MDD	AR	IC	RankIC	ICIR	MDD
Alpha158	0.0808 $\pm$ 0.0009	0.0260 $\pm$ 0.0009	0.0373 $\pm$ 0.0011	0.1663 $\pm$ 0.0041	-0.1161 $\pm$ 0.0047	0.0737 $\pm$ 0.0025	0.0199 $\pm$ 0.0007	0.0254 $\pm$ 0.0007	0.1640 $\pm$ 0.0054	-0.1424 $\pm$ 0.0116	-0.0325 $\pm$ 0.0055	-0.0027 $\pm$ 0.0024	-0.0001 $\pm$ 0.0025	-0.0227 $\pm$ 0.0208	-0.4084 $\pm$ 0.0852
Alpha360	0.0704 $\pm$ 0.0033	0.0231 $\pm$ 0.0017	0.0325 $\pm$ 0.0005	0.1582 $\pm$ 0.0138	-0.1022 $\pm$ 0.0099	0.0703 $\pm$ 0.0007	0.0206 $\pm$ 0.0013	0.0263 $\pm$ 0.0007	0.1676 $\pm$ 0.0113	-0.1113 $\pm$ 0.0050	0.0160 $\pm$ 0.0060	-0.0044 $\pm$ 0.0021	-0.0058 $\pm$ 0.0009	-0.0497 $\pm$ 0.0248	-0.3162 $\pm$ 0.0688
Automated Factor Mining
GP-Learn	0.0815 $\pm$ 0.0013	0.0185 $\pm$ 0.0016	0.0271 $\pm$ 0.0018	0.1181 $\pm$ 0.0093	-0.1132 $\pm$ 0.0106	0.0965 $\pm$ 0.0016	0.0259 $\pm$ 0.0001	0.0368 $\pm$ 0.0002	0.1918 $\pm$ 0.0002	-0.1727 $\pm$ 0.0194	-0.1301 $\pm$ 0.0104	-0.0016 $\pm$ 0.0015	0.0028 $\pm$ 0.0011	-0.0120 $\pm$ 0.0118	-0.5591 $\pm$ 0.0838

Table 1: Backtest performance across CSI 300, CSI 1000, and S&P 500. Each cell reports mean

\pm

standard deviation over five random seeds. Within each dataset and metric, the best result is shown in bold , and the second-best result is underlined .

Figure 2: Cumulative excess return curves across CSI 300, CSI 1000, and S&P 500.

	CSI 300					S&P 500
Setting	AR%	IC	RankIC	ICIR	MDD%	AR%	IC	RankIC	ICIR	MDD%
AlphaRLM (8B)	12.14	0.026	0.042	0.159	-6.15	6.94	0.002	0.003	0.016	-14.84
w/o DAPO	9.66	0.025	0.034	0.141	-7.58	3.14	-0.001	0.004	-0.004	-30.72
Non-financial text w/o DAPO	5.24	0.013	0.035	0.115	-12.99	1.78	-0.003	0.004	-0.022	-26.46
Non-financial text w/ DAPO	2.13	0.009	0.031	0.068	-17.59	2.90	-0.008	-0.002	-0.063	-28.72
Linear reward	9.68	0.031	0.042	0.184	-15.60	0.77	-0.002	0.004	-0.028	-19.88

Table 2: Ablation results for AlphaRLM-8B on CSI 300 and S&P 500.

Figure 3: Matched factor-count comparison of AlphaRLM-8B with Alpha158 and Alpha360 across three markets. Each pair uses 158 or 360 factors.

Figure 4: Factor-level TMB distributions on CSI 300 and S&P 500. The horizontal axis is shown on a symmetric log scale to accommodate zero-valued factors.

Factor 1: Quantile-Clipped Trading Activity per Price Movement

If(Gt(

volume, Quantile(

volume, 20, 0.95)),

Quantile($volume, 20, 0.95),

If(Lt(

volume, Quantile(

volume, 20, 0.05)),

Quantile(

volume, 20, 0.05),

volume))

/ Abs(Delta($close, 1))

Literature input: Patzelt and Bouchaud (2018) .

Table 3: Illustrative AlphaRLM (8B) factors with verified literature inputs.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: DAPO training dynamics for Qwen3-8B on CSI 300 and S&P 500.

Asset	Period	Trading days
CSI 300	Training: 2015-01 to 2019-12	1219
	Validation: 2020-01 to 2020-12	243
	Testing: 2021-01 to 2024-12	968
CSI 1000	Training: 2015-01 to 2019-12	1219
	Validation: 2020-01 to 2020-12	243
	Testing: 2021-01 to 2024-12	968

Appendix

Table 4: Market-data splits for downstream XGBoost modeling. S&P 500 trading-day counts follow the U.S. equity trading calendar.

Market	Factor pool	Count	AR	IC	RankIC	ICIR	MDD
CSI 300	Alpha158	158	8.08	0.0260	0.0373	0.1663	-11.61
	AlphaRLM-top158	158	11.95	0.0239	0.0433	0.1478	-8.22
	Alpha360	360	7.04	0.0231	0.0325	0.1582	-10.22
	AlphaRLM-top360	360	12.65	0.0296	0.0470	0.1840	-7.18
CSI 1000	Alpha158	158	7.37	0.0199	0.0254	0.1640	-14.24
	AlphaRLM-top158	158	11.95	0.0327	0.0384	0.2568	-10.51

Appendix

Table 5: Complete results at matched factor counts. Generated subsets are derived from AlphaRLM-8B. AR and MDD are reported in percentages; fixed-library baselines are reproduced from Table 1 .

Factor Libraries
	CSI 300					CSI 1000					S&P 500
Method	Mean	SD	Median	P95	$P_{0.8}$	Mean	SD	Median	P95	$P_{0.8}$	Mean	SD	Median	P95	$P_{0.8}$
Alpha158	0.2638	0.2316	0.1924	0.7225	2.36%	0.2651	0.2299	0.1976	0.7215	2.45%	0.1680	0.2224	0.0742	0.6902	2.85%
Alpha360	0.4887	0.3176	0.5111	0.9555	22.75%	0.4985	0.3139	0.5337	0.9549	23.03%	0.4747	0.3568	0.5295	0.9681	25.52%
Automated Factor Mining
GP-Learn	0.2819	0.2461	0.1898	0.8223	5.29%	0.2641	0.2333	0.1832	0.7388	3.47%	0.1793	0.2157	0.0830	0.7062	2.89%

Appendix

Table 6: Pairwise absolute Pearson correlations within the final factor pools on the 2021–2024 test period. For each factor pair, correlation is computed by pooling all jointly valid stock–day observations over the full test period.

P_{0.8}

denotes the proportion of factor pairs satisfying

|r|\geq 0.8

. Lower values generally indicate less linear redundancy.

Pool	Exact uniq.	Structural uniq.	Repeated factors	Max cluster	Exact library overlap	Structural library overlap
AlphaRLM-8B A-share	99.9%	95.9%	7.2%	5	0.000%	0.104%
AlphaRLM-8B S&P 500	99.5%	89.1%	18.4%	10	0.500%	0.337%
AlphaRLM-14B A-share	99.7%	80.7%	28.1%	22	0.100%	0.124%
AlphaRLM-14B S&P 500	99.9%	92.9%	11.9%	9	0.000%	0.000%

Appendix

Table 7: Within-pool structural diversity and overlap with the standard Alpha158/Alpha360 libraries. Exact uniqueness is the proportion of distinct exact canonical expressions. Structural uniqueness is the proportion of distinct parameter-normalized skeletons. Repeated-factor ratio denotes the fraction of factors belonging to skeleton clusters of size greater than one. Library structural overlap is computed over unique AlphaRLM skeletons.

Market	Factor count	AR (%)	IC	RankIC	ICIR	MDD (%)
CSI 300	158	11.9469	0.023905	0.043301	0.147774	-8.2226
	360	12.6458	0.029587	0.046999	0.183973	-7.1806
	1,000	12.1375	0.025554	0.042442	0.158593	-6.1504
	1,500	11.1957	0.026342	0.043941	0.159020	-7.4845
	2,000	9.0978	0.025492	0.043189	0.155702	-7.4613
CSI 1000	158	11.9513	0.032682	0.038445	0.256770	-10.5107

Appendix

Table 8: Sensitivity to the number of selected factors for AlphaRLM-8B.

Factor Libraries
	CSI 300					CSI 1000					S&P 500
Method	AR	IC	RankIC	ICIR	MDD	AR	IC	RankIC	ICIR	MDD	AR	IC	RankIC	ICIR	MDD
Alpha158	0.0624	0.0152	0.0333	0.0891	-0.1357	0.0734	0.0010	0.0164	0.0074	-0.1500	-0.1358	-0.0013	0.0017	-0.0092	-0.6318
Alpha360	0.0175	0.0082	0.0209	0.0377	-0.2501	-0.0259	0.0065	0.0303	0.0489	-0.2241	-0.1864	-0.0173	-0.0191	-0.1343	-0.8209
Automated Factor Mining
GP-Learn	0.0150	0.0049	0.0068	0.0280	-0.2109	0.0870	0.0166	0.0365	0.1208	-0.1761	-0.0758	-0.0042	0.0040	-0.0247	-0.4063

Appendix

Table 9: Backtest performance using Ridge regression as the downstream predictor. The factor pools and evaluation protocol are held fixed relative to the main experiments, with only the return predictor replaced. Within each dataset and metric, the best result is shown in bold , and the second-best result is underlined .

Quantitative Finance

Quantitative finance uses mathematical models, statistics, and machine learning to value assets, measure risk, and guide trading and portfolio decisions.

See all 191 papers

From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery
May 14, 2026Lingzhe Zhang, Tong Jia, Yunpeng Zhai +5RL for Language Models Alpha Factor Discovery
Peking University, Beijing, China · Alibaba Group, China · Nanjing University, Nanjing, China +1
AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining
Jul 29, 2026Jingyang Yi, Jian Yang, Yifei Jin +2Automated Algorithm Discovery Alpha Factor Discovery
1X-Tech · 2Xtech-PandaAI-Waton Joint Lab · 3Monash University +2
AlphaLogics: A Market Logic-Driven Multi-Agent System for Scalable and Interpretable Alpha Factor Generation
Mar 10, 2026Zhangyuhua Weng, Shengli Zhang, Taotao Wang +1Alpha Factor Discovery Quantitative Finance
College of Electronic and Information Engineering, Shenzhen University

May 14, 2026J/K move · Enter open · S save

From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery

Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Zixuan Xie+4

Peking University, Beijing, China · Alibaba Group, China · Nanjing University, Nanjing, China+1

RL for Language Models Alpha Factor Discovery Large Language Model-Guided Optimization RL Fine-Tuning Algorithmic Trading

Open paper AskPDF arXiv

Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals. Recent LLM-based methods have shown promise in automating factor generation, but most of them still rely on prompt-level generation--evaluation--feedback loops for iterative optimization. As the loop becomes longer, repeatedly appended historical candidates and feedback can cause context explosion, increase inference cost, dilute useful information, and introduce feedback drift. Moreover, these methods often depend on very large LLMs whose stable generation preferences may lead to structurally similar expressions, redundant candidates, and search stagnation. To address these limitations, we propose QuantEvolver, a self-evolving alpha factor discovery framework based on reinforcement fine-tuning. Instead of accumulating feedback in the prompt, QuantEvolver converts executable quantitative evaluation into policy updates, enabling a Miner LLM to internalize historical optimization experience through parameter learning. Specifically, QuantEvolver constructs high-quality seed factors, builds diverse seed--time-window training tasks, generates executable Factor DSL expressions, evaluates them through Regime Backtest, and optimizes the Miner LLM with Diversity-Complementarity Reward. During training, high-quality factors are continuously accumulated in a Mined Factor Database, which serves as the final discovered factor library. Extensive experiments on three realistic market benchmarks demonstrate the effectiveness of QuantEvolver, which consistently improves the primary evaluation metric of each task over existing LLM-based alpha factor discovery baselines, produces higher-quality and more complementary factor pools.

Abstract

Quantitative Finance

Explore similar work

From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery

AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining

AlphaLogics: A Market Logic-Driven Multi-Agent System for Scalable and Interpretable Alpha Factor Generation