{"paper_id":"2406.02969","title":"Filtered not Mixed: Stochastic Filtering-Based Online Gating for Mixture of Large Language Models","authors":["Raeid Saqur","Anastasis Kratsios","Florian Krach","Yannick Limmer","Jacob-Junqi Tian","John Willes","Blanka Horvath","Frank Rudzicz"],"paper_date":"2024-06-05","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2406.02969v2","page_url":"https://thequant.space/flowcharts/filtered-not-mixed-stochastic-filtering-based-online-gating/","scores":{"empirical_rigor":7,"math_complexity":8.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes MoE-F, a stochastic filtering-based gating mechanism to combine multiple pre-trained LLMs for online time-series prediction.","quote":"We propose MoE-F – a formalized mechanism for combining N pre-trained Large Language Models (LLMs) for online time-series prediction","summary":"We propose MoE-F - a formalized mechanism for combining N pre-trained Large Language Models (LLMs) for online time-series prediction by adaptively forecasting the best weighting of LLM predictions at every time step. Our mechanism leverages the conditional information in each expert’s running…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The target process Y evolves according to a hidden Markov model with a hidden state w that follows a continuous-time Markov chain.","type":"model","stated":"explicit","quote":"... Assume that Y· evolves according to the dynamics in Eq. (4). We assume that w· is a hidden Markov process with values in the standard basis"},{"assumption":"The input path x(t) is once continuously differentiable and the intensity matrix Q is uniformly bounded.","type":"statistical","stated":"explicit","quote":"Assumptions 2.1 (Regularity Conditions). The path x· : [0, ∞) →Rd is once continuously differentiable and there is a constant C > 0 such that, for each t ≥0 and i, j = 1, . . . , d, (Qt)i,j ≤C."}]},"datasets":{"status":"available","items":[{"name":"NIFTY","provider":null,"kind":"text-news","access":"public","period":"2010-01-06 to 2020-09-21","universe":"US equities stock market","quote":"... is a processed and curated daily news headlines dataset for the stock (US Equities) market price movement prediction task."},{"name":"ETTh1","provider":null,"kind":"benchmark","access":"public","period":null,"universe":null,"quote":"time-series forecasting (LTSF) using the ETTh1, ETTh2 [84] datasets with concurrent specialized LTSF models."},{"name":"ETTh2","provider":null,"kind":"benchmark","access":"public","period":null,"universe":null,"quote":"time-series forecasting (LTSF) using the ETTh1, ETTh2 [84] datasets with concurrent specialized LTSF models."}],"known":[{"id":"twitter","name":"Twitter / X","provider":"X Corp.","kind":"text-news","access":"public","quote":"Other stock price movement prediction from social dataset include what is referred to as ACL18 (or, ‘acl18’) in this paper is essentially the StockNet [80] dataset which comprises of stock tweets of 88 stock tickers from 9 financial market industries from Twitter over two years (from 2014-2015) aligned with their corresponding historical price data.","mentions":1,"used_in":1},{"id":"yahoo-finance","name":"Yahoo Finance","provider":"Yahoo","kind":"market-prices","access":"public","quote":"The compilation includes headlines from The Wall Street Journal and Reuters News, as well as market data of the $SPY index from Yahoo Finance.","mentions":1,"used_in":1}],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R1","candidates":{"front_matter":"n/a","evidence_card":"no","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"forecast-accuracy","findings":[{"finding":"MoE-F achieves a 17% absolute and 48.5% relative improvement in F1 score over the next best individual LLM on the Financial Market Movement task.","kind":"performance","quote":"MoE-F attains a remarkable 17% absolute and 48.5% relative F1 measure improvement over the next best performing individual LLM expert predicting short-horizon market movement"},{"finding":"In Table 2, MoE-F shows a 43% F1 score compared to 52% for the best individual expert, confirming theoretical guarantees.","kind":"performance","quote":"the overall performance on the FMM task is lower than from earlier: 43% compared to 52% (Table 2). This aligns with our proposed theoretical performance guarantee relative to experts."},{"finding":"MoE-F consistently outperforms base Llama-2 and Llama-3 models across multiple datasets, achieving up to 0.43 F1 score.","kind":"performance","quote":"Table 3: Performance of Llama-2-7b-chat and Llama-3-8B-Instruct"}],"reported_performance":null,"out_of_sample":{"method":"holdout","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":18,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":1,"extracted_at":"2026-11-14T18:08:03Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}