Source-linked AI summary
CauProphet: Teaching Agents to Develop Causal Cognition like Humans
Kaiwen Wu, Chaochao Lu
TL;DR
Agents lack a way to accumulate and refine causal understanding across experiences, limiting a capability needed to distinguish observed associations from action effects. CauProphet combines retained causal knowledge with current evidence and updates it as observations arrive. Its evaluations report transfer to later decisions, including improved financial research and prediction accuracy in unfolding real-world events.
Problem
How agents can accumulate causal understanding across experiences remains open, despite the need to distinguish observed associations from the effects of actions.
Source: §1 · p. 3Method
CauProphet combines retained causal explanations with current task evidence to form a decision-guiding Causal Belief, then updates the belief and retained knowledge as observations arrive.
Source: §1 · p. 3Results
CauProphet progressively transfers learned causal knowledge to later tasks, with a 25.4% relative gain over an agent with text memory in mean correlation between predicted and actual rankings of next-day stock returns.
Source: §1 · pp. 3–4Takeaways & Limitations
The results support developing agents whose causal understanding extends beyond individual experiences and informs new decisions.
Source: §5 · p. 12When evidence is insufficient, CauProphet relies more on the base model’s prior knowledge; experiments cover few stages, leaving longer-term adaptation in changing real-world settings for future study.
Source: Limitations · p. 12
Abstract
from arXiv · showhide
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. The project repository is available at https://github.com/OpenCausaLab/CauProphet.
1. Introduction
The introduction identifies a gap in how agents accumulate causal understanding across experiences and introduces CauProphet to retain and apply causal knowledge. Evaluations report improved financial research and prediction accuracy, alongside tests of causal analysis and action selection.
1. Introduction
- Prior work addresses causal knowledge construction or uses causal knowledge to guide behavior, leaving open how agents accumulate understanding across experiences.The introduction contrasts this with human learning, which supports predictions and decisions without recalling every shaping experience.Source: §1 · p. 3
- CauProphet combines retained causal explanations with task evidence to form a Causal Belief that guides decisions, then updates both belief and retained knowledge as observations arrive.Its retained explanations capture recurring factors, their effects, and the conditions under which relationships hold.Source: §1 · pp. 3–4
- 25.4% relative gain over an agent with text memory in mean correlation between predicted and actual rankings of next-day stock returns; CauProphet outperforms all baselines.Each financial task begins in a fresh context.Source: §1 · pp. 3–4
- Prediction accuracy increases from 21.3% for a baseline agent to 28.6% for CauProphet as real-world events unfold.Further tests cover causal analysis and action selection in scientific and everyday applications.Source: §1 · p. 4
2. Related Work
The related work reviews agent memory, causal discovery, and causal reasoning for behavior. It identifies lasting abstract causal understanding learned from experience as an open area.
2. Related Work
- Agent systems combine reasoning, tools, and memory, but how agents develop causal understanding across experiences remains open.The cited memory mechanisms support information retention and access beyond the current context.Source: §2 · p. 4
- Causal discovery and analysis methods focus on structures, effect estimates, and analysis procedures; CauProphet focuses on retaining findings and their conditions for later tasks.Reusing findings requires retaining both what relationship was learned and under which conditions it holds.Source: §2 · p. 4
- Prior systems apply causal knowledge to planning, multimodal reasoning, program revision, and execution repair, while lasting abstract causal cognition from experience remains underexplored.Source: §2 · p. 5
3. Methods
CauProphet combines retained causal knowledge with current evidence to guide task decisions, revises its belief as evidence changes, and retains selected findings for later use. Its methods represent causal relationships with conditions and evidence, compare alternative explanations, and update knowledge across tasks.
3.1. Overview
- CauProphet combines retained relationships and their conditions with current evidence to form a Causal Belief that describes influences and possible action effects, guiding decisions.It treats tasks as partially observable and uses Causal Scout to compare explanations and suggest discriminating observations.Source: §3.1 · p. 5
3.2. Causal Understanding across Experiences
- Causal Cognition retains explanations of recurring factors, mechanisms, conditions, time scales, evidence, and unresolved questions in graph or written form.The agent retrieves relevant knowledge using the task goal and circumstances, then maps factors to current variables and checks conditions against evidence.Source: §3.2 · p. 5
- The same retained knowledge can inform different task beliefs; the resulting belief guides actions, and its graph can support causal effect estimation.Source: Figure 2 · p. 6, §3.2 · p. 6
- CauProphet updates confidence in retained relationships from supporting and conflicting claims, assuming claims are independent given whether the relationship holds.The Bayesian update uses prior and updated relationship probabilities and claim likelihood ratios.Source: §3.2 · p. 6
- For updates, supporting claims add confidence-based weights and contradicting claims subtract them; repeated proposals do not accumulate as independent support.Graph retrieval follows relationships with positive log-odds and ranks them by updated probability and relevance to current conditions.Source: §3.2 · p. 6
3.3. Revising Causal Understanding with Evidence
- When explanations fit the current belief, Causal Scout enumerates alternatives under constraints on required or forbidden links, incompatible hypotheses, cycles, and temporal order.It checks candidates against available statistical tests, ranks them by confidence and simplicity, and returns evidence and distinguishing observations.Source: §3.3 · p. 7
- As new evidence arrives, the model checks observations against predictions, conditions, and time periods, revising relationships or assumptions and using suitable data to estimate effects and assess reliability.Analysis results return to Causal Belief and can prompt revised explanations or further analyses.Source: §3.3 · p. 7
- The agent retains selected findings with their mechanisms, conditions, and evidence; findings under different conditions or time scales remain separate.Written explanations are revised directly, and earlier versions and reasons for change are retained for time-appropriate retrieval.Source: §3.3 · p. 7
4. Experiments
Across chronological financial studies, CauProphet leads the baselines and improves through repeated formula search; further tests examine causal analysis, intervention choices, scientific data analysis, and everyday tasks.
4.1. Settings
- Experiments use DeepSeek-V4-Pro, with ReAct and OpenCode as baselines; chronological financial tasks begin in fresh contexts with shared candidate and model-call limits.Financial studies also include text-memory ReAct; the tabular study includes model-only references.Source: §4.1 · pp. 7–8
4.2. Developing Causal Cognition through Experience
- CauProphet achieves the highest mean RankIC across two chronological financial studies, outperforming ReAct, OpenCode, and text-memory ReAct.Earlier feedback also updates retained explanations and informs later formula revisions.Source: §4.2 · p. 8
- Causal Cognition’s held-out score rises from 4.213 to 4.850 in Round 4 and ends at 4.820, while text memory remains at 4.191 across five rounds.Causal Cognition uses retained comparisons to continue testing and revising formulas; text memory selects the same final formula each round.Source: §4.2 · p. 8
- Across 24 FutureSim questions, CauProphet achieves the highest BSS and Top-1 accuracy before events resolve.The replay yields 195 question-day predictions per system, including 171 before resolution.Source: §4.2 · p. 8, Figure 5 · p. 10, Table 2 · p. 10
4.3. Foundational Causal Abilities
- CauProphet achieves the lowest mean effect relative error across three SCMs and gets effect direction correct in all nine evaluations.The systems comprise two nonlinear and one linear model, each with 800 observations.Source: §4.3 · p. 8, Figure 6 · p. 10
- CauProphet has higher mean survival and score than ReAct and OpenCode in all three CausalGame scenarios, winning every run in both trap scenarios.The complete system also exceeds both ablations, including one retaining causal computation but removing dedicated causal capabilities.Source: §4.3 · pp. 9–10, Figure 7 · p. 11
4.4. Causal Reasoning across Diverse Domains
- On 12 BLADE tasks, CauProphet matches the best model selection, achieves the highest transformation match rates, and improves variable match over ReAct.The clearest gains concern preparing data as observations and comparisons that answer causal questions.Source: §4.4 · p. 11, Table 3 · p. 11
- Across 23 completed executions of 16 selected AppWorld tasks, CauProphet passes 173 of 174 checks.Tasks span music, messaging, expenses, files, and shopping.Source: Table 4 · p. 11, §4.4 · p. 12
5. Conclusion
CauProphet connects retained causal relationships and their conditions with current evidence to guide decisions, and findings can inform later tasks. Experiments support causal knowledge extending beyond individual experiences.
5. Conclusion
- CauProphet applies retained causal relationships and their conditions to new situations, combining them with current evidence to guide decisions.Findings from tasks can contribute to causal knowledge available for later use.Source: §5 · p. 12
- Experiments show that earlier findings guide subsequent research and that CauProphet makes more accurate predictions while events unfold.The conclusion also notes evaluations of foundational causal abilities and applications to scientific analysis and everyday activities.Source: §5 · p. 12
- The results support developing agents whose causal understanding extends beyond individual experiences and informs new decisions.Source: §5 · p. 12
Limitations
When evidence is insufficient to test causal relationships, CauProphet relies more heavily on the base model’s prior knowledge; the section also illustrates its workflow through case studies, including a Labubu purchase decision.
Limitations
- When evidence cannot test causal relationships, CauProphet relies more heavily on the base model’s prior knowledge; richer observations and action feedback could help assess accumulated knowledge reliability.The experiments cover a limited number of stages; longer studies in changing real-world settings could examine how causal understanding grows and adapts.Source: Limitations · p. 12
A. Case Studies
- The case studies connect user questions to evidence, recorded Causal Beliefs, and knowledge retained in Causal Cognition, with each conversation starting from one request.The Labubu and mathematics conversations were run using CauProphet with Claude Opus 5.5.Source: A. Case Studies · p. 16
A.1. Labubu Prices Are Falling: Buy Now or Wait?
- For the Labubu purchase, CauProphet distinguishes group revenue from demand for a particular figure and proposes that blind-box buyers resell unwanted regular figures, increasing resale supply.It recommends buying at about RMB 160 or below, or waiting seven to ten days at RMB 160–200; it saves the proposed mechanism and distinction for later retrieval.Source: §A.1 · p. 16
A.2. Does AI Improve Mathematics Learning and Exam Scores?
Using real classroom data, CauProphet distinguishes AI-assisted practice from independent exam performance: practice scores rise, but the no-AI exam shows no clear improvement, and the agent retains this lesson.
A.2. Does AI Improve Mathematics Learning and Exam Scores?
- The supplied classroom study compares answer-giving AI, hint-providing AI with teacher-prepared materials, and no AI, then tests all groups without AI.The data and study description let CauProphet distinguish performance with assistance from independent learning.Source: §A.2 · p. 16
- On the no-AI exam, differences versus students without AI were −5.4 points for answer-giving AI and −0.4 points for hint-providing AI.The respective 95% intervals were [−9.8, −1.0] and [−3.1, 2.2]; practice scores increased by 13.7 points with answer-giving AI and 36.1 points with hint-providing AI, after rescaling to 100 points.Source: Figure 9 · p. 18, §A.2 · p. 19
- The analysis excludes honors classes and uses 2,848 student–lesson records from 825 students in 44 classes, accounting for prior grades and classroom factors.Uncertainty is calculated at the class level; practice scores are not adjusted for when estimating exam effects because AI use already affects them.Source: §A.2 · p. 19
- CauProphet records five variables and five directed edges, marks the practice-to-exam link as a hypothesis, and saves the distinction between assisted practice gains and independent exam gains.Its search finds no directly applicable causal relationship; the persistent graph’s existing concepts and relationships retain their recorded state.Source: §A.2 · p. 19
B. Method and Implementation Details
Causal Cognition retains recurring factors and proposed relationships with their mechanisms, conditions, time scales, and evidence, while written explanations can preserve mechanisms even without a graph.
B.1. Representing Causal Knowledge
- Causal Cognition retains recurring factors, proposed relationships, mechanisms, conditions, time scales, and supporting or conflicting evidence.For graph knowledge, a source link connects a relationship to the experience that motivated it.Source: §B.1 · p. 19
- In the Labubu example, the retained supply–demand explanation distinguishes supply from demand when interpreting a resale-price decline.The Causal Belief also includes variables and observations needed to interpret current reports.Source: §B.1 · p. 19
- Written explanations can retain a mechanism and its conditions even when the agent has not constructed a graph.Source: §B.1 · p. 19
B.2. Applying Knowledge to a New Task
CauProphet checks whether retained explanations fit a new situation, revises beliefs and retained knowledge as evidence changes, and compares alternative explanations against supplied constraints.
B.2. Applying Knowledge to a New Task
- For a new task, CauProphet maps recurring factors to current variables and checks whether retained conditions and time scales still fit before using knowledge to guide a reply or investigation.It records the selected written-explanation version and its account of how that explanation applies to the task.Source: §B.2 · p. 19
B.3. Revising Knowledge with Evidence
- New observations can support, contradict, or expose missing conditions in an explanation, prompting revisions to the current belief and retained knowledge.Findings under different conditions or time scales remain separate; written explanations preserve reasons for changes and earlier versions, while retrieval selects the latest version available at the task’s time.Source: §B.3 · pp. 19–20
B.4. Comparing Alternative Explanations
- Causal Scout checks candidate explanations against relationship, temporal-order, and statistical-evidence constraints, then ranks alternatives by confidence and simplicity.It returns supporting evidence, counterevidence, and observations that could distinguish alternatives, informing which explanation to investigate or adopt.Source: §B.4 · p. 20
C. Experimental Details and Additional Analyses
The financial studies test whether retained knowledge transfers across chronological and repeated formula-search tasks, with fresh task contexts and held-out evaluation. They report the protocols and metrics used to compare workflows, but these passages do not provide comparative outcome values.
C.1.1. Chronological Studies
- The chronological studies search CSI300 formulas for next-day returns across six half-years per study, with the submitted formula fixed before feedback from the next half-year.The price study permits price and volume inputs in 2014–2016; the volume study permits only volume and historical transformations in 2017–2019.Source: §C.1.1 · p. 20
- Common dates across four workflows are used within each half-year, and the six periods are weighted equally for reporting.Tables 5–7 cover per-period results, research costs, and submitted formulas; RankIC is displayed multiplied by 100.Source: §C.1.1 · p. 20, Table 5 · p. 21
C.1.2. Iterative Formula Search
- The fixed-data study runs MA, ROC, and BETA starting formulas through five rounds, comparing Causal Cognition with text memory under the same base agent.Candidate feedback uses 2017–2018 data, while held-out scoring uses 2019 data that is not returned to the agent.Source: §C.1.2 · p. 21
- The fixed-data study reports held-out scores and search effort by round, plus RankIC by starting formula and round.IC averages daily IC over the scoring window; RankICIR relates mean RankIC to its variation and is not annualized.Source: §C.1.2 · p. 21, Table 8 · p. 22
C.1.3. Evolution of Causal Cognition
These analyses describe how retained explanations and comparisons inform later financial searches, then specify evaluation settings for causal action selection, scientific analysis, forecasting, and application tasks. The supplied passages provide protocols and selected process observations, not comparative result values for these evaluations.
C.1.3. Evolution of Causal Cognition
- In chronological searches, CauProphet retains a five-day price-reversal formula while refining its account of tested alternatives and where the signal worked.For volume, it records period-varying signal strength and later compares alternative baselines, including a 10-day baseline for 2019H2; fixed-data traces revisit prior comparisons and add new outcomes.Source: §C.1.3 · pp. 22–23, Table 9 · p. 23
C.2. CausalGame
- CausalGame tests three environmental scenarios in repeated runs, comparing CauProphet with ReAct, OpenCode, and two causal-capability ablations.The ablations remove either the belief, cognition, and scout components while retaining causal tools, or all dedicated causal capabilities.Source: §C.2 · p. 23
C.3. BLADE
- BLADE compares three systems on the same 12 dataset–research-question pairs using model, variable, value-transform, and graph-transform matches.Reported scores equally average task-level match rates across the 12 tasks.Source: §C.3 · p. 23
C.4. FutureSim
- FutureSim replays 24 forecasting questions with date-appropriate news access and scores predictions against resolved answers using semantic matching.It reports pre-resolution averages by question and across questions, as well as a final-snapshot score based on one resolution-day prediction per question.Source: §C.4 · pp. 24–25
C.5. AppWorld
- AppWorld evaluates selected tasks across five application domains from prescribed states and fresh sessions, checking task completion and application-state changes.The reported set contains 16 selected tasks and 23 executions, including training tasks and three Test-Challenge tasks used for research calibration.Source: §C.5 · p. 25
C.6. Structural Causal Models
The structural causal model evaluation compares recovered structures, adjustment sets, and intervention-effect estimates against known answers. The accompanying prompt and tool instructions specify how CauProphet establishes and revises task-local beliefs, applies retained knowledge, investigates uncertainty, and handles shared research tasks.
C.6. Structural Causal Models
- Three SCMs with 800 observations each are evaluated on recovered directed edges, adjustment sets, and estimated intervention effects, with three evaluations per construction.The true graph and intervention effects are withheld from methods; model-only references answer in one call without tools.Source: §C.6 · p. 25, Table 14 · p. 27
D.1. Establishing a Working Explanation
- The prompts require a working explanation that records variables, justified or hypothesized edges, uncertainty, and rationales for final directed edges.Causal Belief is task-local working state, while Causal Cognition is long-term memory across tasks and conversations.Source: §D.1 · pp. 27–28
D.2. Using and Revising Retained Knowledge
- Retained relations and notes are checked against their measurement, context, lag, mechanism, evidence, and conditions before application to a task belief.Useful findings can revise stored explanations or add selected relations; existing relations should be updated rather than duplicated.Source: §D.2 · pp. 28–29
D.3. Investigating and Answering
- Investigation is matched to available evidence: with suitable data, tools can investigate effects, while without data the agent reasons qualitatively and states needed evidence.The instructions require explicit assumptions and uncertainty, and caution against definitive results when key assumptions are unverified.Source: §D.3 · p. 29
- Causal Scout compares candidate DAGs using d-separation and returns competing graphs, direction constraints, rejection reasons, and discriminating probes.It does not prove hidden confounders or rewrite the active Causal Belief.Source: §D.3 · p. 29
D.4. Shared Research and Memory Instructions
- Shared research instructions treat observations as evidence rather than instructions and make submitted formulas immutable; memory use and maintenance remain preferences.Agents are asked to consult relevant stored notes, retain useful observations, and revise them when evidence changes.Source: §D.4 · p. 30