Research
F2Agent paper proposes a multimodal fusion architecture for LLM trading agents, reporting over 20% average annualised-return improvement against 16 baseline methods across six stocks and crypto assets
Submitted 6 August 2026, the paper argues existing multimodal trading agents fail to capture cross-modal dependencies between price, text and other signal types and remain fragile to market noise. F2Agent's fix is a modality-aware fusion mechanism plus a noise-robust consistency regularizer, tested on six assets including GOOG and TSLA.
F2Agent sits in a specific and increasingly crowded sub-field of the LLM-trading-agent research this site has tracked through August: agents that combine multiple input types, price series, news text, filings, social sentiment, rather than reasoning over a single data stream. The paper's core claim is that prior multimodal approaches bolt different modalities together without properly modelling how they depend on each other, and that this weak fusion is what makes those agents brittle when the market gets noisy. F2Agent's answer is a hierarchy of specialised agents that extract signals per modality, combined through a fusion mechanism designed to be adaptive to which modality is actually informative at a given moment, backed by a regularisation term meant to stop noise in one input from corrupting the combined signal.
The headline numbers, a reported 20%-plus average improvement in annualised return over 16 baseline methods, with individual figures of 120% on GOOG and 148% on TSLA, are the kind of backtest results that invite scepticism on sight. Six assets is a narrow test universe, and large single-name return figures on a two-year AI-melt-up stock like TSLA are exactly the pattern this site has flagged before as a lookahead or overfitting risk unless the paper's evaluation window and cost assumptions are scrutinised directly, something a 32-page methods-heavy submission from earlier this month has not yet had time to receive from the wider research community.
What is more durable than the specific return figures is the direction the paper is pointing in: the field has largely moved past asking whether an LLM can trade at all, covered exhaustively in this site's research pieces since June, and into a second-generation question of how to combine heterogeneous inputs without one noisy signal degrading the rest. That mirrors a pattern familiar from other applied-ML domains, once a baseline capability is established, the interesting work shifts to robustness and input fusion rather than raw capability claims.
For readers tracking this literature rather than deploying it, F2Agent is worth filing alongside this month's other benchmark papers, KTD-Fin's masking-based alpha attribution and Backtrader-Bench's real-vs-mined-question gap, as another entry making the same underlying point from a different angle: headline backtest returns on LLM trading agents keep arriving faster than independent replication of them does, and the gap between the two is where the genuine research risk sits.
Read the original: arXiv - F2Agent paper proposes a multimodal fusion architecture for LLM trading agents, reporting over 20% average annualised-return improvement against 16 baseline methods across six stocks and crypto assets. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.