CauProphet: Teaching Agents to Develop Causal Cognition like Humans

Kaiwen Wu; Chaochao Lu

cs.AIOct 23, 2026

CauProphet: Teaching Agents to Develop Causal Cognition like Humans

Authors: Kaiwen Wu, Chaochao Lu

Organizations: Shanghai Jiao Tong University · Shanghai AI Laboratory

Abstract

Large language model (LLM) agents increasingly support long-horizon tasks by leveraging external memory and past experience. However, existing agents mainly retrieve records or summaries, lacking the ability to form and refine causal understanding of the world. Inspired by human causal cognition, we introduce CauProphet, an agent that develops and updates causal knowledge within its Causal Cognition. For each task, CauProphet combines relevant knowledge in Causal Cognition with current evidence to construct a Causal Belief, a structured belief capturing causal factors and the effects of possible actions. It uses this belief for decision-making and continuously updates both the belief and Causal Cognition as new observations emerge, enabling causal knowledge that persists beyond individual experiences. Evaluations demonstrate that CauProphet progressively acquires causal knowledge and transfers it to future decision-making. In financial research, CauProphet leverages prior discoveries to improve downstream analyses and outperforms all baselines in mean rank correlation with next-day returns, achieving a 25.4% improvement over a baseline agent. It also improves prediction accuracy from 21.3% to 28.6% in dynamic real-world settings. Further experiments validate its ability to perform reliable causal analysis and action selection across scientific and everyday tasks, demonstrating the potential of causal experience learning for next-generation LLM agents.

Figures & tables

Figure 1: Overview of CauProphet . CauProphet maintains explicit causal hypotheses and uses causal skills and tools to test and revise them, keeping its reasoning traceable to evidence and assumptions.

Figure 2: Architecture of CauProphet . Causal Cognition retains abstract causal knowledge across tasks, relevant parts of which are combined with the current task’s evidence to form Causal Belief.

Figure 3: Financial research across successive periods. CauProphet achieves the highest mean score across the two studies. Each point scores a committed formula on the next half-year.

	Next-period RankIC $\times 100$ $\uparrow$
Method	Price	Volume	Mean
ReAct	6.119	2.792	4.455
OpenCode	6.901	1.488	4.195
ReAct + text memory	5.172	3.041	4.106
CauProphet	7.418	2.881	5.149

Table 1: Chronological factor research, RankIC

\times 100

. We use 694/722 common dates for price/volume, average within each half-year, then weight periods and studies equally.

Figure 4: Knowledge reuse across five rounds of formula search. Mean held-out RankIC

\times 100

for Causal Cognition and text memory; higher is better.

Figure 6: Causal analysis on three SCMs. CauProphet achieves the lowest mean relative error in effect estimation across all three systems. Outward is better:

F_{1}

scores and Backdoor/Sign success rates use

[0,1]

; SHD and fractional relative error

e

use

1/(1+e)

. Curves show transformed means; raw means and SDs are in Table 13 .

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: An explanation develops through interaction. (A) An ordinary question leads CauProphet to create a task-local Causal Belief and explain the observed case count. (B) The user’s clarification about symptom-targeted testing supplies a condition that changes the interpretation of test positivity. The panels show excerpts from the interaction.

Figure 10: From a revised belief to retained causal understanding. (C) The task-local graph includes the testing-selection mechanism introduced during the conversation. (D) Causal Cognition retains the selected relationship, an explanation of its applicable condition, and a link to the source episode.

Study	Period	Dates	ReAct	OpenCode	ReAct + text	CauProphet
Price	2014H1	117	6.759	6.759	6.759	6.664
	2014H2	125	6.621	5.639	7.127	7.127
	2015H1	111	0.017	4.888	6.000	6.000
	2015H2	117	7.420	7.420	8.013	8.013
	2016H1	101	8.449	8.449	8.196	8.449
	2016H2	123	7.446	8.253	-5.064	8.253

Appendix

Table 4: Per-period chronological AlphaBench-derived results. Values are signed mean next-day-return RankIC

\times 100

on dates common to all four workflows. H1 and H2 denote the first and second half of the year. Each study has one chronological trajectory per workflow.

Study	Method	Tokens	Initialization tokens	Model calls	Candidates	Reads	Writes
Price	ReAct	58,741	0	29	36	0	0
	OpenCode	95,913	0	31	36	0	0
	ReAct + text memory	116,299	0	41	36	11	6
	CauProphet	211,247	28,220	43	42	7	8
Volume	ReAct	59,832	0	29	36	0	0
	OpenCode	91,498	0	30	36	0	0

Appendix

Table 5: Full recorded cost of chronological financial research, including the historical task, optional post-submission maintenance, and recorded failed calls. Reads and writes count memory-tool calls, including unsuccessful attempts. Successful CauProphet writes are 7 of 8 in the price study and 8 of 9 in the volume study; text-memory writes all succeed.

Study	Period	Submitted formula
Price	2014H1	Ref( $close, 5) /$ close
	2014H2	Ref( $close, 5) /$ close
	2015H1	Ref( $close, 5) /$ close
	2015H2	Ref( $close, 5) /$ close
	2016H1	Ref( $close, 5) /$ close
	2016H2	Ref( $close,5)/$ close

Appendix

Table 6: CauProphet ’s immutable submissions in the chronological studies. Ref selects a past value; Mean averages a historical window; Log takes a logarithm. Close denotes closing price and volume denotes trading volume.

Memory	Round	RankIC $\times 100$	RankICIR	IC	Model calls	Candidates valid/attempted
Causal Cognition	1	4.213	0.267	0.014	27	17/18
	2	4.282	0.271	0.015	25	15/18
	3	4.484	0.285	0.014	23	18/18
	4	4.850	0.312	0.021	27	17/18
	5	4.820	0.305	0.017	27	15/18
Text memory	1	4.191	0.267	0.013	22	15/18

Appendix

Table 7: Causal Cognition and text memory on AlphaBench: mean held-out scores and total search effort per round. Higher is better for the three score columns. Valid/attempted counts exclude the initial seed. Each memory condition follows one continuing five-round sequence covering three starting formulas.

Round	Starting formula	Causal Cognition	Text memory
1	MA	4.191	4.191
1	ROC	4.167	4.191
1	BETA	4.282	4.191
2	MA	4.282	4.191
2	ROC	4.282	4.191
2	BETA	4.282	4.191

Appendix

Table 8: Held-out RankIC

\times 100

by starting formula and round (higher is better). MA is a moving-average ratio, ROC a reversal ratio, and BETA a normalized price slope. The starting formulas share the same search objective, market, and date windows.

Antenna trap
Scenario / configuration	Survival (%)	Score (%)	Wins
CauProphet	86.4 $\pm$ 7.0	81.5 $\pm$ 4.9	3/3
ReAct	55.9 $\pm$ 14.7	60.1 $\pm$ 10.3	0/3
OpenCode	68.4 $\pm$ 7.6	68.9 $\pm$ 5.3	1/3
w/o causal components	78.7 $\pm$ 8.6	76.1 $\pm$ 6.0	2/3
w/o causal capabilities	73.4 $\pm$ 10.7	72.4 $\pm$ 7.5	2/3

Appendix

Table 9: CausalGame aggregate results, with means

\pm

sample standard deviations over three runs per configuration and scenario. Wins count victories out of three runs.

Dataset	Research question focus
affairs	Children and extramarital affairs
amtl	Tooth loss in humans and other primates
boxes	Age, culture, and reliance on majority preference
caschools	Student–teacher ratio and academic performance
crofoot	Group size, location, and intergroup contest outcomes
fish	Factors associated with fish caught per hour

Appendix

Table 10: The 12 BLADE dataset–research-question pairs. Dataset names are the benchmark’s public identifiers.

Metric	CauProphet	ReAct	OpenCode
Time-weighted BSS	0.080	0.049	0.035
Time-weighted Top-1 (%)	28.6	21.3	22.0
Time-weighted truth probability (%)	14.7	13.2	12.6
Final-snapshot BSS	0.085	0.092	0.006
Final-snapshot Top-1 (%)	29.2	29.2	29.2
Final-snapshot truth probability (%)	16.3	16.9	15.0

Appendix

Table 11: FutureSim metrics across the same 24 questions for all three systems. BSS, Top-1 accuracy, and truth probability are higher-is-better measures. Counts report scoring coverage.

Official task ID	Task summary	Checks per execution
Training tasks
82e2fac_1	Find the most-liked song in a playlist.	2/2
82e2fac_2	Find the least-played song in the library.	2/2
29caf6f_1	Extract movie recommendations from notes and reply to a message.	8/8
afc0fce_1	Find recent coworker payments, then like and comment on them.	9/9
6104387_1	Export a complete, deduplicated music library and delete the Spotify account.	10/10; 10/10; 10/10; 10/10

Appendix

Table 12: Selected AppWorld tasks and completed executions. Multiple entries in a row list separate completed executions of that task in chronological order.

		Structure		Adjustment		Intervention effect
SCM construction	Method	Edge $F_{1}$ (%) $\uparrow$	SHD $\downarrow$	Adj. $F_{1}$ (%) $\uparrow$	Backdoor	RelErr (%) $\downarrow$	Sign
Synthetic 5-node nonlinear	CauProphet	100.0 $\pm$ 0.0	0.000 $\pm$ 0.000	100.0 $\pm$ 0.0	3/3	1.1 $\pm$ 1.4	3/3
	ReAct	94.9 $\pm$ 4.4	0.667 $\pm$ 0.577	100.0 $\pm$ 0.0	3/3	124.3 $\pm$ 104.5	1/3
	OpenCode	89.7 $\pm$ 5.5	1.333 $\pm$ 0.577	88.9 $\pm$ 19.2	2/3	82.0 $\pm$ 80.6	2/3
	DeepSeek- V4-Pro	64.5 $\pm$ 13.6	3.667 $\pm$ 1.155	88.9 $\pm$ 19.2	2/3	70.3 $\pm$ 28.5	2/3
	GPT-6 Sol	95.2 $\pm$ 8.2	0.667 $\pm$ 1.155	100.0 $\pm$ 0.0	3/3	23.6 $\pm$ 10.8	3/3

Appendix

Table 13: Structure, adjustment, and intervention effects on three SCMs. Means

\pm

sample SD over three evaluations; Backdoor/Sign count valid adjustment sets/correct effect directions. SHD counts graph edits; RelErr is effect relative error. Agent rows use DeepSeek-V4-Pro; model-only rows use one call without tools.

Causal Reasoning in Language Models

Causal reasoning in language models studies how models identify cause-and-effect structure, predict interventions, and answer counterfactual questions.

See all 102 papers

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists
May 25, 2026Junlin Yang, Dylan Zhang, Xiangchen Song +7Structural Causal Models LLM Agent Evaluation
Tsinghua University · University of Illinois Urbana-Champaign · Carnegie Mellon University +2
Synergizing Large Language Models and Theory for Human-like Causal Reasoning
May 13, 2025Yanxi Zhang, Xin Cong, Zhong Zhang +3Causal Reasoning in Language Models Causal Reasoning
Center for Data Science, AAIS, Peking University · Wangxuan Institute of Computer Technology, Peking University · Tsinghua University +1
Causal Discovery in the Era of Agents
Jun 22, 2026Yujia Zheng, Vishal Verma, Mantej Gill +3AI Agents for Scientific Discovery Causal Discovery
1Carnegie Mellon University. · 2Mohamed bin Zayed University of Artificial Intelligence.

May 25, 2026J/K move · Enter open · S save

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists

Junlin Yang, Dylan Zhang, Xiangchen Song, Qirun Dai+6

Tsinghua University · University of Illinois Urbana-Champaign · Carnegie Mellon University+2

Structural Causal Models LLM Agent Evaluation Causal Discovery AI Agent Benchmarks Causal Reasoning in Language Models

Open paper AskPDF arXiv

We introduce CausaLab, a scalable environment for evaluating interactive causal discovery by LLM agents. Unlike prior evaluations, CausaLab evaluates both whether an agent can solve a problem using causal evidence and whether its answer is grounded in a faithful recovered causal mechanism. Each episode places an agent in a synthetic laboratory: it receives prior measurement records, intervenes on a manipulator crystal, and predicts the resonance frequency of a held-out reactor crystal governed by the same mechanism. The hidden data-generating process is a randomly sampled structural causal model (SCM), so success requires recovering both a causal graph and structural equations rather than recalling prior knowledge. Experiments show a persistent gap between prediction and mechanism recovery: in the purely observational 6-node setting, GPT-5.2-high reaches 92% task accuracy but only 0.471 all-edge

F_1

. Mixed observation-intervention strategies improve structural fidelity, while pure intervention remains difficult even for strong agents. We identify premature stopping as a major weakness and show that consistency verification mitigates it. CausaLab therefore separates predictive success from causal understanding and exposes current LLM agents' limits as experimental causal reasoners.

Abstract

Causal Reasoning in Language Models

Explore similar work

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists

Synergizing Large Language Models and Theory for Human-like Causal Reasoning

Causal Discovery in the Era of Agents