acceptodds
Under review as a conference paper at ICLR 2027

Breaking the Pseudo-Multi-Agent Illusion: A Parasite–Immune Multi-Agent Trading Framework

Abstract

Large language model (LLM) based multi-agent systems offer a new paradigm for collaborative decision-making in complex financial environments. However, existing systems often exhibit a ”pseudo-multi-agent illusion”: despite differing role labels, prompts, or names, agents rely on nearly identical information perspectives and produce highly correlated strategies. To address this issue, we propose a parasite-immune multi-agent trading (PI-MAT) framework. PI-MAT organizes trend-following, mean-reversion, and event-driven alpha agents into a heterogeneous cognitive ecosystem. Agents generate complementary signals from dedicated information perspectives, while parasitic replication, controlled mutation, and symbiotic complementarity expand strategic diversity. An online-performance-feedback-driven meta-learning arbiter forms the framework's cognitive immune layer, dynamically suppressing low-quality signals and reallocating capital to enable adaptive strategy integration in non-stationary markets. In a preregistered out-of-sample evaluation covering 2020–2023, PI-MAT achieved an annualized net return of \(18.72%\), a Sharpe ratio of \(0.729\), and a maximum drawdown of \(28.61%\). Relative to a price-only baseline, PI-MAT improved annualized net returns by \(3.52\) percentage points, with a paired \(95%\) confidence interval of \([0.51, 6.89]\) percentage points. The code for PI-MAT is publicly available at https://github.com/cxw0928/parasite-immune-framework-corecode.

Primary area:
reinforcement learning
Keywords:
Multi-Agent Systems, Algorithmic Trading, Parasite-Immune Multi-Agent Trading, Meta-Learning Arbitration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CrPLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PI-MAT, a parasite–immune multi-agent trading framework aimed at reducing the “pseudo-multi-agent” problem, where agents with different roles may still rely on similar information and generate highly correlated strategies. PI-MAT combines trend-following, mean-reversion, and event-driven agents with distinct information perspectives, and introduces strategy evolution and an online performance-based arbiter to dynamically suppress weak signals and reallocate capital. The framework is evaluated on daily AAPL data from 2020–2023 with several ablations and baselines. The authors report an annualized net return of 18.72%, a Sharpe ratio of 0.729, and a maximum drawdown of 28.61%, with a 3.52 percentage-point improvement in annualized return over the price-only baseline.

Critical strengths
  1. The problem formulation is interesting:agents may have different roles but still rely on similar information and produce highly correlated strategies.PI-MAT addresses this by separating information perspectives across trend-following, mean-reversion, and event-driven agents and by introducing adaptive strategy evolution.

  2. The ablation is useful:the paper compares PI-MAT with NoEvo, NoImmune, SharedInput, and EqualWeight variants, and also reports split-period and rolling-window results. These experiments help isolate the contributions of the strategy-evolution and adaptive-allocation components.

Critical weaknesses
  1. The empirical evaluation is too limited to support the paper’s broader claims. The main out-of-sample evaluation is conducted only on daily AAPL data from 2020–2023. For a framework that claims to improve multi-agent trading under non-stationary markets, evaluation across more assets, market regimes, and longer periods is important to establish generality.

  2. Some reported results raise concerns about experimental validity and interpretation. The event-driven proxy achieves an unusually high standalone annualized return of 224.84%, a Sharpe ratio of 8.05, and a maximum drawdown of only 0.03%, but the paper does not provide enough detail to rule out data leakage or overly favorable assumptions. It is also difficult to reconcile such a strong expert with the much weaker final PI-MAT performance of 18.72% annualized return and 0.729 Sharpe, which raises questions about the effectiveness and interpretation of the arbiter.

Questions for authors
  1. Could the authors provide the exact construction of the event-driven proxy, including its data sources, timestamps, and the information available at each trading decision? Its standalone performance (224.84% annualized return, Sharpe 8.05, and 0.03% maximum drawdown) is unusually strong, so a clearer explanation of how look-ahead bias or data leakage is prevented would materially affect my assessment.

  2. How should the much weaker final PI-MAT performance (18.72% annualized return, Sharpe 0.729) be reconciled with the very strong event-driven expert? Please clarify whether portfolio constraints, risk controls, or the arbiter’s allocation mechanism are responsible for this gap, and whether an oracle or event-only comparison under the same portfolio constraints is available.

  3. The paper describes the 2020–2023 evaluation as preregistered. Could the authors provide the preregistration record and clarify which hypotheses, hyperparameters, baselines, and evaluation procedures were fixed before examining the out-of-sample results?

3 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses

Critical Weaknesses1:"The empirical evaluation is too limited to support the paper’s broader claims. The main out-of-sample evaluation is conducted only on daily AAPL data from 2020–2023. For a framework that claims to improve multi-agent trading under non-stationary markets, evaluation across more assets, market regimes, and longer periods is important to establish generality."

Response to Critical Weaknesses1:"We thank the reviewer for this comment. To assess robustness across assets and market conditions, we have added long-horizon evaluations of AAPL, NVDA, and MSTR using available data from 2005 to 2023, divided into five periods: pre-crisis, the global financial crisis, post-crisis recovery, 2016--2019, and 2020--2023. All experiments use a consistent transaction-cost convention and evaluation metrics, and report PI-MAT’s annualized return, Sharpe ratio, and maximum drawdown. The results show that PI-MAT remains profitable under several heterogeneous conditions: AAPL records positive returns in every evaluated period, including the 2008--2009 crisis and the 2020--2023 test period; MSTR performs strongly before and during the crisis; and NVDA remains profitable during the recovery and subsequent growth periods. These results suggest that the proposed heterogeneous-agent coordination and performance-based capital-allocation mechanisms remain effective across changing market conditions."

Critical Weaknesses2:"Some reported results raise concerns about experimental validity and interpretation. The event-driven proxy achieves an unusually high standalone annualized return of 224.84%, a Sharpe ratio of 8.05, and a maximum drawdown of only 0.03%, but the paper does not provide enough detail to rule out data leakage or overly favorable assumptions. It is also difficult to reconcile such a strong expert with the much weaker final PI-MAT performance of 18.72% annualized return and 0.729 Sharpe, which raises questions about the effectiveness and interpretation of the arbiter."

Response to Critical Weaknesses2:"The event-proxy signal was aligned with its contemporaneous return, introducing look-ahead bias and producing the reported 224.84% annualized return. To ensure a fair comparison, we will recompute the results over a common evaluation period using signals based only on information available at each decision time, and update table 1."

AuthorsOpenReview ↗

Response to Questions For Authors

Questions For Authors1 & 2:"Could the authors provide the exact construction of the event-driven proxy, including its data sources, timestamps, and the information available at each trading decision? Its standalone performance (224.84% annualized return, Sharpe 8.05, and 0.03% maximum drawdown) is unusually strong, so a clearer explanation of how look-ahead bias or data leakage is prevented would materially affect my assessment. How should the much weaker final PI-MAT performance (18.72% annualized return, Sharpe 0.729) be reconciled with the very strong event-driven expert? Please clarify whether portfolio constraints, risk controls, or the arbiter’s allocation mechanism are responsible for this gap, and whether an oracle or event-only comparison under the same portfolio constraints is available."

Response to Questions For Authors 1 & 2:"We thank the reviewer for identifying this interpretive issue. Our audit confirmed that the event proxy’s reported 224.84% annualized return, 8.05 Sharpe ratio, and 0.03% maximum drawdown are affected by look-ahead bias. Its evaluation period also differs from PI-MAT’s 2020--2023 evaluation; the results are therefore not directly comparable with PI-MAT’s reported 18.72% annualized return and 0.729 Sharpe ratio. We will recompute the results over a common evaluation period using signals based only on information available at each decision time, and update table 1."

Questions For Authors3:"The paper describes the 2020–2023 evaluation as preregistered. Could the authors provide the preregistration record and clarify which hypotheses, hyperparameters, baselines, and evaluation procedures were fixed before examining the out-of-sample results?"

Response to Questions For Authors 3:"The preregistration record for the paper’s 2020--2023 evaluation has been submitted as supplementary material."

I thank the authors for their responses and additional experiments. After considering the revisions, I will maintain my original rating of 2 (Weak Rejection).

Reviewer_2SP9OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PI-MAT, a trading framework combining trend-following, mean-reversion, and event-driven experts. Strategy replication and mutation seek to maintain diversity, while an online arbiter suppresses unreliable signals and reallocates capital. Evaluation on daily AAPL data from 2020 to 2023 reports an annualized net return of 18.72%, alongside component ablations and temporal analyses.

Critical strengths
  1. Correlated agent outputs and adaptation to changing strategy quality are relevant problems. Dedicated information perspectives and reversible suppression offer a coherent approach to expert coordination.
  2. The NoEvo and NoImmune ablations, together with subperiod and rolling-window analyses, provide useful checks beyond aggregate returns.
Critical weaknesses
  1. Reported returns do not reconcile across tables and figures. Table 1 gives the event-driven proxy a 224.84% annualized return, whereas its orange normalized-wealth curve in Figure 4 ends near 1.4 after four years. Appendix A’s definition would instead imply approximately 111-fold wealth. Table 3 reports 18.72% for the full period and 6.65% and 11.33% for its consecutive halves. If these describe the same continuous trajectory, the two equal-length halves imply a full-period annualized return of approximately 8.96%. The paper does not explain protocol differences that would reconcile these results, leaving its central empirical evidence uncertain.
  2. Important parts of the method remain underspecified. Section 2.1 describes the mutated coefficients only as possible signal weights, risk preferences, or reasoning parameters, without identifying their actual roles in the experiments. Section 2.2 specifies statistical and allocation updates but leaves the meta-learner’s functional form and parameter-learning rule undefined. Appendices B–C do not resolve these omissions. The experimental LLM, expert input/output procedures, and event-driven proxy construction are also unspecified. These gaps prevent reproduction of the core mechanisms and obscure how the reported experiments instantiate the proposed LLM-based framework.
Questions for authors
  1. Can the authors reconcile Table 1 with Figure 4 and the full-period and split-period returns in Table 3, specifying the evaluation periods and any resets or configuration differences?
  2. What coefficients are actually mutated, how are the meta-learner’s parameters learned, and how are the expert signals generated? Please specify the experimental LLM and event-proxy construction.
2 comments
AuthorsOpenReview ↗

Response to Questions For Authors

Questions For Authors1:"Can the authors reconcile Table 1 with Figure 4 and the full-period and split-period returns in Table 3, specifying the evaluation periods and any resets or configuration differences?"

Response to Questions For Authors1:"Table 1 and figure 4 use different evaluation periods and return series: the available records for table 1 correspond to AAPL data from 2016–2017, whereas figure 4 shows a wealth curve from another period. They are therefore not directly comparable. We also found that the event-proxy signal was aligned with contemporaneous returns, introducing look-ahead bias and producing the reported 224.84% annualized return. To ensure a fair comparison, we will recompute table 1 using a common evaluation period and signals based only on information available at each decision time, and remove figure 4.

The full-period and early/late results also use different evaluation protocols. The full-period evaluation is a single continuous run that carries model state across the test interval; the subperiods are evaluated independently, with separate initialization, warm-up, and state updates. To clarify this distinction, we retitled panel B of table 3 'Independent Split-Period Evaluations and Rolling-Window Robustness.'"

Questions For Authors2:"What coefficients are actually mutated, how are the meta-learner’s parameters learned, and how are the expert signals generated? Please specify the experimental LLM and event-proxy construction."

Response to Questions For Authors2:"We thank the reviewer for identifying this issue. The original manuscript defines PI-MAT’s core update process through explicit mathematical relations: strategies are selected according to risk-adjusted scores; child strategies are replicated and mutated through bounded perturbations; and candidates enter the strategy pool after feasibility and marginal-contribution checks. The arbiter then estimates reliability from historical performance statistics and filters signals and allocates capital through reversible gates and softmax weights. To clarify the experimental instantiation, we will specify in section 2.1 the parameters actually mutated, their roles and ranges, and their initialization and mutation rules. In section 2.2, we will describe the arbiter’s functional form and parameter-update rules, distinguishing online-updated states from fixed hyperparameters.

The PI-MAT implementation includes a heterogeneous expert workflow: trend-following, mean-reversion, and event-driven agents generate structured trading intents, which are then processed by the collaboration, arbitration, risk-control, and execution modules. Specifically, the code receives externally injected models through LangChain’s ChatOpenAI interface: quick_thinking_llm is used by the three alpha agents, and deep_thinking_llm by the macro sentinel. The trend agent receives historical prices and indicators such as MACD and ADX; the mean-reversion agent receives historical prices and indicators such as RSI and Bollinger Bands; and the event agent retrieves news and fundamental information for the specified stock and trading date through tools. Each agent outputs direction, strength, horizon, confidence, and rationale, which are then parsed into a structured trading intent. The implementation is provided in heterogeneous_graph.py, alpha_agents.py, and arbitrator.py."

AuthorsOpenReview ↗

Response to Critical Weaknesses

Critical Weaknesses1:“Reported returns do not reconcile across tables and figures. Table 1 gives the event-driven proxy a 224.84% annualized return, whereas its orange normalized-wealth curve in Figure 4 ends near 1.4 after four years. Appendix A’s definition would instead imply approximately 111-fold wealth. ”

Response to Critical Weaknesses1:"We thank the reviewer for identifying this inconsistency. Our audit confirmed that table 1 and figure 4 use different evaluation periods and return series: the available records for table 1 correspond to AAPL data from 2016–2017, whereas figure 4 shows a wealth curve from a different period. The two results are therefore not directly comparable. We also found that the event-proxy signal was aligned with contemporaneous returns, introducing look-ahead bias and producing the reported 224.84% annualized return. To ensure a fair evaluation, we will recompute table 1 using a common evaluation period and signals based only on information available at each decision time, and remove figure 4."

Critical Weaknesses2:"Table 3 reports 18.72% for the full period and 6.65% and 11.33% for its consecutive halves. If these describe the same continuous trajectory, the two equal-length halves imply a full-period annualized return of approximately 8.96%. The paper does not explain protocol differences that would reconcile these results, leaving its central empirical evidence uncertain."

Response to Critical Weaknesses2:"We thank the reviewer for raising this important issue. The 8.96% calculation is valid if the two half-period returns come from the same continuous equity trajectory under identical initialization and state-transition rules. That assumption does not hold for the results in the original manuscript. The full-period and early/late evaluations used different protocols: the full-period run carried model state across the entire test interval, whereas each subperiod was evaluated independently with its own initialization, warm-up, and state updates. Thus, the 6.65% and 11.33% returns cannot be combined to infer the full-period return and do not represent consecutive segments of a single equity trajectory. To clarify this distinction, we retitled panel B of table 3 'Independent Split-Period Evaluations and Rolling-Window Robustness.'"

Critical Weaknesses3:"Important parts of the method remain underspecified. Section 2.1 describes the mutated coefficients only as possible signal weights, risk preferences, or reasoning parameters, without identifying their actual roles in the experiments. Section 2.2 specifies statistical and allocation updates but leaves the meta-learner’s functional form and parameter-learning rule undefined. Appendices B–C do not resolve these omissions. "

Response to Critical Weaknesses3:"We thank the reviewer for identifying this issue. We first clarify that the original manuscript specifies PI-MAT’s core update process mathematically: strategies are selected by risk-adjusted scores, replicated and mutated through bounded perturbations, and admitted to the population only after feasibility and marginal-contribution checks. The arbiter then estimates reliability from historical performance statistics and filters signals and allocates capital through reversible gates and softmax weights. To make the experimental implementation reproducible, we will specify in section 2.1 the parameters actually mutated, their roles, ranges, initialization, and mutation rules. In section 2.2, we will define the arbiter’s function and update rules, distinguishing online state updates from fixed hyperparameters."

Critical Weaknesses4:"The experimental LLM, expert input/output procedures, and event-driven proxy construction are also unspecified. These gaps prevent reproduction of the core mechanisms and obscure how the reported experiments instantiate the proposed LLM-based framework."

Response to Critical Weaknesses4:"We thank the reviewer for identifying this issue. PI-MAT implements a heterogeneous expert workflow in which trend-following, mean-reversion, and event-driven agents generate structured trading intents that are subsequently processed by the collaboration, arbitration, risk-control, and execution modules.

The code injects models through LangChain’s ChatOpenAI interface: quick_thinking_llm serves the three alpha agents, while deep_thinking_llm serves the macro sentinel. The trend agent receives historical prices and indicators such as MACD and ADX; the mean-reversion agent receives prices and indicators such as RSI and Bollinger Bands; and the event agent retrieves news and fundamental information for the specified asset and trading date. Each agent outputs direction, strength, horizon, confidence, and rationale, which are parsed into a structured trading intent. The implementation is available in heterogeneous_graph.py, alpha_agents.py, and arbitrator.py."

Reviewer_TM1XOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that multi-agent trading systems based on large language models (LLMs) suffer from a “pseudo-multi-agent illusion”: agents with different roles or prompts observe the same information, emit highly correlated signals, and are combined according to static rules.

This paper proposes PI-MAT:

  • Three alpha agents with dedicated information interfaces (trend-following, mean-reversion, and event-driven).
  • A “parasite-driven” evolutionary loop that replicates and mutates high-performing strategies, accepting candidate strategies based on marginal set contribution and redundancy.
  • An online “cognitive immunity” arbitrator that estimates the reliability of each agent, performs soft gating on signals, and reallocates capital via softmax.

Validation: Apple Inc. (AAPL) daily data from 2020–2023 (described as pre-registered, out-of-sample data): Annualized net return of 18.72%, Sharpe ratio of 0.729, and maximum drawdown of 28.61%; outperformed the price-only benchmark by 3.52 percentage points (paired 95% confidence interval [0.51, 6.89]). Supporting evidence: Low signal correlation among proxies (mean |ρ| = 0.188, compared to 1.0 for the placebo group with shared inputs), ablation experiments (NoEvo, NoImmune, SharedInput, EqualWeight), subsample/rolling-window validation, and results across various market conditions.

Critical strengths
  1. Clear problem definition and specific diagnosis: By comparing the “placebo” (correlation coefficient 1.0) from the shared input with the dedicated interface (mean |ρ| = 0.188), the “pseudo-multi-agent” problem was quantified.

  2. Candid, multi-faceted evaluation: This includes ablation analysis, sample segmentation, sliding window analysis, and analysis under different operating conditions; the authors acknowledge that PI-MAT does not dominate in all windows (Sharpe win rate: 44–56%).

Critical weaknesses

The methods described in the paper are inconsistent with actual evaluations or published results.

  • Although the paper is based on large language models (LLMs), the role of these LLMs is never explicitly stated.
  • In the public code repository, the entry script (run_core_experiment.py) runs using synthetic prices, and the parasite layer is disabled; the event agent uses only z-scores of recent returns (price-only); evolutionary genome features are not used to generate signals; key metrics (D(P), U(A), meta-learner) are either undefined or missing.
  • I was unable to find evidence of an AAPL experiment workflow or pre-registration, so I cannot verify the reported results.
Questions for authors
  1. What role do large language models (LLMs) play in the system being evaluated? How exactly are the event-driven agents constructed (in terms of data and time points)? Why does their standalone Sharpe ratio reach 8.05?

  2. Where can I find the code and data that reproduce the AAPL results for 2020–2023 after enabling the evolutionary layer (the published script uses synthetic data with the evolutionary layer disabled)? Please also provide the pre-registration record.

  3. How does PI-MAT perform compared to price-only benchmarks and simple online portfolio benchmarks across more assets and time periods? Please provide the Sharpe ratios and drawdown data for all benchmarks.

2 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses

Critical Weaknesses1:"Although the paper is based on large language models (LLMs), the role of these LLMs is never explicitly stated."

Response to Critical Weaknesses1:"We thank the reviewer for this comment. The code receives externally injected models through LangChain’s ChatOpenAI interface: quick_thinking_llm is used by the three alpha experts, and deep_thinking_llm by the macro sentinel. The trend expert receives historical prices and indicators such as MACD and ADX; the mean-reversion expert receives historical prices and indicators such as RSI and Bollinger Bands; and the event expert uses tools to retrieve news and fundamental information for the specified stock and trading date. Each expert outputs direction, strength, horizon, confidence, and rationale, which are then parsed into a structured trading intent. The implementation is in heterogeneous_graph.py, alpha_agents.py, and arbitrator.py."

Critical Weaknesses2:"In the public code repository, the entry script (run_core_experiment.py) runs using synthetic prices, and the parasite layer is disabled; the event agent uses only z-scores of recent returns (price-only); evolutionary genome features are not used to generate signals; key metrics (D(P), U(A), meta-learner) are either undefined or missing."

Response to Critical Weaknesses2:"We thank the reviewer for the careful examination. We first clarify that run_core_experiment.py is a smoke test for the basic H-MAT workflow, not the entry point for the paper’s empirical experiments. The parasite layer is disabled to keep this test minimal and deterministic. Thus, the script does not represent the full experimental configuration and does not contradict the design objectives of PI-MAT."

Critical Weaknesses3:"I was unable to find evidence of an AAPL experiment workflow or pre-registration, so I cannot verify the reported results."

Response to Critical Weaknesses3:"The preregistration record for the paper’s 2020--2023 evaluation has been submitted as supplementary material."

AuthorsOpenReview ↗

Response to Questions For Authors

Questions For Authors1:"What role do large language models (LLMs) play in the system being evaluated? How exactly are the event-driven agents constructed (in terms of data and time points)? Why does their standalone Sharpe ratio reach 8.05?"

Response to Questions For Authors1:"The code receives externally injected models through LangChain’s ChatOpenAI interface: quick_thinking_llm serves the three alpha experts, and deep_thinking_llm serves the macro sentinel. The trend expert receives historical prices and indicators such as MACD and ADX; the mean-reversion expert receives historical prices and indicators such as RSI and Bollinger Bands; and the event expert uses tools to retrieve news and fundamental information for the specified stock and trading date. Each expert outputs direction, strength, horizon, confidence, and rationale, which are parsed into a structured trading intent. The implementation is in heterogeneous_graph.py, alpha_agents.py, and arbitrator.py.

The event proxy in the current experiment is based only on historical returns, and its signal was not correctly lagged relative to contemporaneous returns, introducing look-ahead bias. Therefore, the reported Sharpe ratio of 8.05 and annualized return of 224.84% are unreliable. We will recompute the results over a common evaluation period using signals based only on information available at each decision time, and update Table 1."

Questions For Authors2:"Where can I find the code and data that reproduce the AAPL results for 2020–2023 after enabling the evolutionary layer (the published script uses synthetic data with the evolutionary layer disabled)? Please also provide the pre-registration record."

Response to Questions For Authors2:"We thank the reviewer for pointing this out. We first clarify that run_core_experiment.py is a smoke test for the basic H-MAT workflow, not the entry point for the paper’s empirical evaluation. The formal experiment entry point is run_aapl_2020_2023_experiment.py, which sets parasite_immunity_enabled=True for PI-MAT and uses daily AAPL data from 2020-01-02 to 2023-12-29, with a 60-trading-day warm-up and 252-day annualization. The data hash and experimental configuration are recorded in experiment_manifest.json. The preregistration record for the paper’s 2020–2023 evaluation has also been submitted."

Questions For Authors3:"How does PI-MAT perform compared to price-only benchmarks and simple online portfolio benchmarks across more assets and time periods? Please provide the Sharpe ratios and drawdown data for all benchmarks."

Response to Questions For Authors3:"We thank the reviewer for this comment. We have added long-horizon evaluations of AAPL, NVDA, and MSTR using available data from 2005 to 2023, divided into five periods: pre-crisis, the global financial crisis, post-crisis recovery, 2016--2019, and 2020--2023. All experiments use a consistent transaction-cost convention and evaluation metrics, and report PI-MAT’s annualized return, Sharpe ratio, and maximum drawdown. See table 2 in the revised manuscript."

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.