FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness

# FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness
Source: [https://arxiv.org/html/2606.17642](https://arxiv.org/html/2606.17642)
Pianran Guo1,\*,Pengcheng Zhou2,\*,Yuchen Jian1,Shuhua Chen1

1Beijing University of Posts and Telecommunications,2Queen Mary University of London guopianran@bupt\.edu\.cn,jp2020213547@qmul\.ac\.uk, jianyucheng@bupt\.edu\.cn,lyonia\_orangerica@bupt\.edu\.cn \*Equal contribution

###### Abstract

Financial multimodal reasoning requires agents to coordinate numerical computation, retrieval, visual interpretation, and temporal grounding across heterogeneous evidence sources\. Existing tool\-augmented agents improve execution fidelity, yet remain largely stateless across episodes, repeatedly rediscovering reasoning strategies and failure patterns\. In high\-stakes financial settings, this leads to unreliable tool routing, noisy retrieval, and hallucination\-prone reasoning\. We present FinAcumen, a financial reasoning agent framework centered on selective experience memory for tool\-augmented multimodal reasoning\. FinAcumen accumulates financially grounded reasoning experience from prior trajectories, distilling successful strategies and failure\-derived cautionary rules into a persistent memory bank\. During inference, retrieved experiences condition reasoning only when semantic relevance exceeds a calibrated threshold, while irrelevant memory is explicitly suppressed through a fallback mechanism\. A deterministic financial tool environment further grounds numerical computation, retrieval, visual decoding, and answer verification\.Across four financial multimodal reasoning benchmarks, FinAcumen consistently improves a frozen 8B vision\-language model over finance\-specialized models and approaches leading proprietary general\-purpose models\. Further analysis shows that selective experience activation improves reasoning reliability under retrieval uncertainty\. Our code is anonymously available at[https://anonymous\.4open\.science/r/FinAcumen](https://anonymous.4open.science/r/FinAcumen)

FinAcumen: Financial Multimodal Reasoning via Self\-Evolving Experience Memory Harness

Pianran Guo1,\*, Pengcheng Zhou2,\*, Yuchen Jian1, Shuhua Chen11Beijing University of Posts and Telecommunications,2Queen Mary University of Londonguopianran@bupt\.edu\.cn,jp2020213547@qmul\.ac\.uk,jianyucheng@bupt\.edu\.cn,lyonia\_orangerica@bupt\.edu\.cn\*Equal contribution

## 1Introduction

![Refer to caption](https://arxiv.org/html/2606.17642v1/x1.png)Figure 1:Example items from the four financial multimodal benchmarks evaluated in this work\.Financial multimodal reasoning requires agents to coordinate heterogeneous evidence from text, tables, charts, and time\-indexed records\. Unlike conventional QA, this setting jointly demands visual interpretation, numerical computation, retrieval, and temporal grounding across multiple interacting modalities\. Figure[1](https://arxiv.org/html/2606.17642#S1.F1)illustrates example items from the four benchmarks we evaluate in this work\. Recent benchmarks expose complementary aspects of this challenge, including SEC\-grounded quantity localization, chart\-centric numerical reasoning, temporal\-aware multimodal retrieval, and hallucination\-sensitive financial analysis\(Koncel\-Kedziorskiet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib11); Tanget al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib6); Zhuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib4); Luoet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib5)\)\.

Despite rapid progress in multimodal large language models, reliable financial reasoning remains unresolved\. General\-purpose models lack financially grounded reasoning strategies under complex multimodal settings, while finance\-specialized models often exhibit limited robustness outside their tuning distributions\(Liuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib1); Huanget al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib2); Caillautet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib3)\)\. Recent evidence shows that even models performing competitively on distribution\-near tasks can fail sharply on chart\-dense and retrieval\-intensive settings, suggesting that static fine\-tuning alone does not provide adaptive reasoning policies for heterogeneous modality compositions\(Daiet al\.,[2026](https://arxiv.org/html/2606.17642#bib.bib19); Denget al\.,[2026](https://arxiv.org/html/2606.17642#bib.bib25)\)\. Performance remains particularly constrained in high\-difficulty settings, where benchmark leaders still exhibit substantial degradation under retrieval ambiguity, multi\-step numerical reasoning, and hallucination\-sensitive evaluation\(Tanget al\.,[2026](https://arxiv.org/html/2606.17642#bib.bib21),[2025](https://arxiv.org/html/2606.17642#bib.bib6); Luoet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib5); Zhuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib4)\)\.

In practical financial analysis workflows, human experts rely not only on domain knowledge, but also on accumulated experience about when to retrieve evidence, invoke tools, cross\-check modalities, and verify intermediate conclusions\. Tool\-augmented paradigms such as ReAct externalize computation and perception, yet remain largely stateless across episodes and repeatedly rediscover reasoning procedures\(Yaoet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib13)\)\. Existing memory\-based paradigms partially address this limitation by storing prior reflections or trajectories, but they often conflate successful and failed experiences, retrieve weakly related guidance, and expand without controlling retrieval quality\(Shinnet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib14); Zhaoet al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib17); Ouyanget al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib7); Zhanget al\.,[2026](https://arxiv.org/html/2606.17642#bib.bib9)\)\. Such behavior is particularly problematic in finance, where irrelevant retrieval can directly degrade reasoning quality and hallucinated evidence may lead to incorrect analytical conclusions\(Shiet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib16); Liuet al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib18)\)\. Benchmarks such as FinTMMBench and FinMME make this issue explicit by penalizing retrieval errors and unsupported reasoning, requiring systems to determine not only what to retrieve, but also whether retrieved guidance should influence inference\(Zhuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib4); Luoet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib5)\)\.

We proposeFinAcumen, a financial multimodal reasoning framework centered on selective experience\-guided inference for tool\-augmented agents\. FinAcumen introduces aFinancial Memory \(FM\)module that accumulates financially grounded reasoning experience from prior trajectories, separating reusable strategies from failure\-derived cautionary rules\. During inference, retrieved experience conditions reasoning only when semantic relevance exceeds a calibrated threshold; otherwise, the model explicitly falls back to its base reasoning policy to suppress noisy memory injection\. A complementaryFinancial Tools \(FT\)environment provides deterministic support for numerical computation, retrieval, visual decoding, and answer verification\.

Our contributions are summarized as follows:

1. 1\.We introduce a selective experience memory framework for financial multimodal agents that separates reusable reasoning strategies from failure\-derived cautionary guidance\.
2. 2\.We propose a retrieval\-conditioned inference mechanism in which memory activates only under sufficient semantic relevance, enabling fallback under retrieval uncertainty\.
3. 3\.Across four financial multimodal reasoning benchmarks, FinAcumen consistently improves a frozen 8B VLM over general\-purpose and finance\-specialized baselines, while approaching leading proprietary models\.

## 2Related Work

#### Financial multimodal reasoning\.

Financial QA evaluation has expanded from text\-only settings to multimodal, temporal\-aware, and hallucination\-sensitive benchmarks\. Representative datasets cover SEC\-grounded quantity localization \(BizBench, where we evaluate the SEC\-NUM subset\), chart\-centric numerical reasoning \(FinMMR\), temporal multimodal retrieval and reasoning \(FinTMMBench\), and broad financial multimodal assessment \(FinMME\)\(Koncel\-Kedziorskiet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib11); Tanget al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib6); Zhuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib4); Luoet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib5)\)\. On the modeling side, finance\-oriented LLMs such as Fin\-R1\(Liuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib1)\), Open\-FinLLMs\(Huanget al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib2)\), and LLM Pro Finance Suite\(Caillautet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib3)\)mainly inject domain knowledge through parameter updates\. In contrast, our formulation in §[3](https://arxiv.org/html/2606.17642#S3)and §[4](https://arxiv.org/html/2606.17642#S4)keeps the backbone frozen and adapts at inference time through a fixed tool suite and retrieved experience, so performance differences are attributable to experience conditioning rather than model re\-training\.

#### Cross\-episode experience memory\.

Research on cross\-episode agent improvement can be viewed along three steps: memory construction, retrieval decision, and bank maintenance\. ReAct\(Yaoet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib13)\)established tool\-interleaved reasoning but remains episode\-local\. Reflexion\(Shinnet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib14)\), ExpeL\(Zhaoet al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib17)\), ReasoningBank\(Ouyanget al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib7)\), and EchoSafe\(Zhanget al\.,[2026](https://arxiv.org/html/2606.17642#bib.bib9)\)move to persistent experience accumulation across episodes\. For retrieval, prior evidence shows that irrelevant context can harmreasoning quality\(Shiet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib16); Liuet al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib18)\), while Self\-RAG\(Asaiet al\.,[2024](https://arxiv.org/html/2606.17642#bib.bib15)\)highlights explicit retrieve\-or\-not decisions\. Relative to these lines, FinAcumen’s FM in §[4\.2](https://arxiv.org/html/2606.17642#S4.SS2)explicitly separates guidance from cautions, activates memory through similarity\-thresholded retrieval with an empty\-return fallback, and applies deduplication\-aware ranking before prompt injection\.

## 3Problem Formulation

Notation\.We consider multimodal financial QA with disjoint train and test splits\. Each instancex=\(q,c,I\)x=\(q,c,I\)comprises questionqq, textual contextcc, and optional imagesII, with gold answery⋆​\(x\)y^\{\\star\}\(x\)scored under benchmark metricμ\\mu\. A frozen VLMπθ\\pi\_\{\\theta\}performs multi\-turn tool\-augmented decoding over a fixed tool suite𝒯\\mathcal\{T\}\(§[4\.4](https://arxiv.org/html/2606.17642#S4.SS4)\)\. A memory bankℳ\\mathcal\{M\}, populated from training trajectories, stores structured experience entries\. At inference, entries are retrieved via embedding similarity and the top\-kkare injected into the prompt \(Algorithm[1](https://arxiv.org/html/2606.17642#alg1)\); a complete symbol table appears in Appendix[A](https://arxiv.org/html/2606.17642#A1)\.

Givenℳ\\mathcal\{M\}populated from𝒟train\\mathcal\{D\}\_\{\\text\{train\}\}trajectories, we seekℳ⋆\\mathcal\{M\}^\{\\star\}that maximizes expected accuracy on𝒟train\\mathcal\{D\}\_\{\\text\{train\}\}:

ℳ⋆∈arg⁡maxℳ⁡𝔼x∼𝒟train​\[μ​\(πθ​\(x;ℳ,𝒯\),y⋆​\(x\)\)\]\.\\mathcal\{M\}^\{\\star\}\\in\\arg\\max\_\{\\mathcal\{M\}\}\\;\\mathbb\{E\}\_\{x\\sim\\mathcal\{D\}\_\{\\text\{train\}\}\}\\Bigl\[\\,\\mu\\bigl\(\\pi\_\{\\theta\}\(x;\\mathcal\{M\},\\mathcal\{T\}\),\\,y^\{\\star\}\(x\)\\bigr\)\\Bigr\]\.\(1\)At test time,ℳ⋆\\mathcal\{M\}^\{\\star\}is frozen and read\-only: queries may activate stored patterns but never write to the bank\. Final performance is reported as𝔼x∼𝒟test​\[μ​\(πθ​\(x;ℳ⋆,𝒯\),y⋆​\(x\)\)\]\\mathbb\{E\}\_\{x\\sim\\mathcal\{D\}\_\{\\text\{test\}\}\}\[\\mu\(\\pi\_\{\\theta\}\(x;\\mathcal\{M\}^\{\\star\},\\mathcal\{T\}\),y^\{\\star\}\(x\)\)\]\.

The discrete optimization in \([1](https://arxiv.org/html/2606.17642#S3.E1)\) is approximated in Section[4](https://arxiv.org/html/2606.17642#S4)through a writer\-retriever loop \(Algorithm[1](https://arxiv.org/html/2606.17642#alg1)\) that constructsℳ\\mathcal\{M\}from graded training trajectories\.

## 4Method

### 4\.1Overview

As defined in Section[3](https://arxiv.org/html/2606.17642#S3), a frozen language modelπθ\\pi\_\{\\theta\}operates over a tool suite𝒯\\mathcal\{T\}to solve financial problems\.FinAcumenextends this formulation with two components, illustrated in Figure[2](https://arxiv.org/html/2606.17642#S4.F2): Financial Memory \(FM\) and Financial Tools \(FT\)\.FMenables the model to accumulate experience from its own solution trajectories and retrieve it as guidance for new problems, growing a bank of generalized strategies and guard rules through multi\-trajectory sampling\.FTprovides a deterministic execution layer that offloads computation, data retrieval, visual decoding, and answer consolidation from the generative process\. The pipeline is shared across accumulation and inference, isolating memory retrieval as the sole conditioned pathway\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x2.png)Figure 2:TheFinAcumenpipeline\. Experience accumulated through multi\-trajectory sampling is stored inFMas generalized strategies and guard rules\. At inference, semantically similar entries are retrieved and provided as in\-context guidance; when none qualify, the model proceeds withFTalone\.
### 4\.2Financial Memory

#### Memory bank structure\.

Each entry inℳ\\mathcal\{M\}is a structured record of one completed problem, containing the original question, its gold answer, and the distilled experience\. The experience consists of two parts: strategies generalized from trajectories that reached the correct answer, and guard rules extracted from those that did not\.

#### Memory consolidation\.

For each problem the model encounters, the framework generates multiple independent solution trajectories fromπθ\\pi\_\{\\theta\}with diverse decoding, scores each against the gold answer, and feeds the scored set to a single summary agent\. Let\{Traj1,…,TrajK\}\\\{\\mathrm\{Traj\}\_\{1\},\\dots,\\mathrm\{Traj\}\_\{K\}\\\}denote the trajectories for problemxxwith gold answery∗y^\{\*\}, and letsk=Score​\(Trajk,y∗\)s\_\{k\}=\\text\{Score\}\(\\mathrm\{Traj\}\_\{k\},y^\{\*\}\)be the correctness score assigned to each\. The summary agent synthesizes two outputs:

\(S,C\)=Summarize​\(\{\(Trajk,sk\)\}k=1K,y∗\),\(S,C\)=\\text\{Summarize\}\\bigl\(\\\{\(\\mathrm\{Traj\}\_\{k\},s\_\{k\}\)\\\}\_\{k=1\}^\{K\},\\;y^\{\*\}\\bigr\),\(2\)whereSScollects strategies from successful trajectories andCCcollects cautions from failed ones\. These are written into a new entry alongside the original question and answer, incrementally expanding the bank for future retrieval\.

#### Memory activation\.

At inference, a queryxxis embedded into the same semantic space as the bank through the shared encoderE​\(⋅\)E\(\\cdot\)\. Its similarity to each entrymmis defined by cosine distance over the resulting vectors:

sim​\(x,m\)=E​\(x\)⋅zm‖E​\(x\)‖​‖zm‖,\\text\{sim\}\(x,m\)=\\frac\{E\(x\)\\cdot z\_\{m\}\}\{\\\|E\(x\)\\\|\\,\\\|z\_\{m\}\\\|\},\(3\)wherezmz\_\{m\}denotes the pre\-stored embedding of entrymm\. Entries passing a calibrated thresholdτ\\tauform the candidate set

ℳx=\{m∈ℳ∣sim​\(x,m\)≥τ\},\\mathcal\{M\}\_\{x\}=\\\{\\,m\\in\\mathcal\{M\}\\mid\\text\{sim\}\(x,m\)\\geq\\tau\\,\\\},\(4\)which is then deduplicated and ranked\. The top\-ranked subsetℳx∗\\mathcal\{M\}\_\{x\}^\{\*\}is rendered as a structured prefix prepended before the problem prompt\. The model produces its final answeryythrough a tool\-augmented rollout

y=\{Rollout​\(πθ,xmem,𝒯\),ℳx∗≠∅Rollout​\(πθ,x,𝒯\),otherwise,y=\\begin\{cases\}\\text\{Rollout\}\\bigl\(\\pi\_\{\\theta\},\\;x\_\{\\text\{mem\}\},\\;\\mathcal\{T\}\\bigr\),&\\mathcal\{M\}\_\{x\}^\{\*\}\\neq\\emptyset\\\\\[4\.0pt\] \\text\{Rollout\}\\bigl\(\\pi\_\{\\theta\},\\;x,\\;\\mathcal\{T\}\\bigr\),&\\text\{otherwise\},\\end\{cases\}\(5\)wherexmemx\_\{\\text\{mem\}\}denotes the prompt with memory prefix prepended\.

### 4\.3Agent Architecture

All LLM\-driven roles inFinAcumenshare the same base model and are distinguished by their prompt\-defined behavior\. During consolidation, the model acts as a sampling agent that generates multiple diverse solution trajectories, and as a summary agent that scores them against the gold answer before synthesizing strategies from successes and cautions from failures\. During inference, an annotator agent rewrites retrieved experiences into directives personalized to the current problem before the solver agent plans and executes tool\-augmented reasoning\. The solver internally decomposes into two phases: planning, where it inventories inputs, judges experience applicability, and formulates a strategy; and execution, where it invokes tools and terminates with the final answer\.

### 4\.4Financial Tools

The tool suite𝒯\\mathcal\{T\}comprises four tools:

𝒯=\{Numerical Reasoning Engine,Grounded Data Retrieval,Visual Structure Decoder,Answer Consolidation Gate\},\\mathcal\{T\}=\\left\\\{\\begin\{gathered\}\\text\{Numerical Reasoning Engine\},\\\\ \\text\{Grounded Data Retrieval\},\\\\ \\text\{Visual Structure Decoder\},\\\\ \\text\{Answer Consolidation Gate\}\\end\{gathered\}\\right\\\},\(6\)covering precise arithmetic, factual data lookup, chart interpretation, and answer verification\. The same tool set and calling schema are shared across all configurations, so that tool mechanics remain identical whether or notFMconditions the prompt\.

#### Numerical Reasoning Engine\.

This tool decouples financial arithmetic from the generative process by executing code in a persistent namespace\. Letσt\\sigma\_\{t\}denote the variable bindings accumulated across consecutive calls within the same agent loop, withσ0\\sigma\_\{0\}containing no bindings:

\(ot,σt\)=Execute​\(codet,σt−1\),\(o\_\{t\},\\sigma\_\{t\}\)=\\textsc\{Execute\}\(\\text\{code\}\_\{t\},\\;\\sigma\_\{t\-1\}\),\(7\)whereoto\_\{t\}is the captured output\. This stateful design enables cascaded operations such as rate compounding and cross\-statement aggregation that stateless code generators cannot reliably chain: without persistent state, the model must reinitialize all quantities in every code block, risking inconsistency across arithmetic steps\. Financial multi\-step reasoning chains intermediate quantities by nature: growth rates build on prior period values, ratios depend on previously computed numerators and denominators\. A fresh execution session per code call severs this dependency chain; the persistent namespace preserves it\.

#### Grounded Data Retrieval\.

This tool resolves factual financial queries against a curated panel of OHLC records, news corpora, and quarterly indicator tables\. Given a symbol, date, and indicator, the lookup returns the corresponding structured entry:

r=Lookup​\(symbol,date,indicator\),r=\\textsc\{Lookup\}\(\\text\{symbol\},\\;\\text\{date\},\\;\\text\{indicator\}\),\(8\)anchoring each numerical claim to a verifiable source row\. All filtering logic executes deterministically within the tool backend, preventing the model from fabricating lookup parameters or results\. Financial quantities such as quarterly revenue or daily closing prices are factual, time\-bound values\. Parametric knowledge, encoded during pre\-training, cannot reliably reproduce them when the query demands a specific date or indicator; consulting a deterministic panel anchors each claim in an externally verifiable source\.

#### Visual Structure Decoder\.

This tool handles chart and table images in two layers\. Before the agent loop begins, all available images are transcribed via OCR and prepended to the context as structured text, providing precise numerical inputs upfront\. During reasoning, if the agent detects that computation results contradict values previously read from images, it may explicitly invoke this tool to re\-parse the source image, producing an updated structured representation that resolves the discrepancy\.

#### Answer Consolidation Gate\.

This tool serves as the sole termination point of the agent loop\. Given the accumulated trajectoryTraj\\mathrm\{Traj\}and the proposed answeraa, it enforces three conditions before emission:

Verify​\(Traj,a\)≡\\displaystyle\\textsc\{Verify\}\(\\mathrm\{Traj\},a\)\\;\\equivsource\-traceable​\(Traj\)\\displaystyle\\text\{source\-traceable\}\(\\mathrm\{Traj\}\)\(9\)∧\\displaystyle\\wedgeunits\-ok​\(Traj\)\\displaystyle\\text\{units\-ok\}\(\\mathrm\{Traj\}\)∧\\displaystyle\\wedgeformat\-valid​\(a\),\\displaystyle\\text\{format\-valid\}\(a\),where source\-traceable checks thataaor its numeric equivalent appears among the captured outputs of Execute calls, the returned entries of Lookup calls, or the OCR\-extracted values in the context; units\-ok verifies that quantity units remain compatible across tool calls without unresolved conversions; and format\-valid ensuresaaconforms to the task instruction\. Because termination is already the final reasoning step, embedding the verification check there costs no additional inference; a separate post\-hoc auditor would require a second traversal of the full trajectory\.

#### Interface withFM\.

FMis the memory bank described in §[4\.2](https://arxiv.org/html/2606.17642#S4.SS2);πθ\\pi\_\{\\theta\}is the frozen language model\.FMoperates strictly at the prompt level and never interleaves with the tool\-calling mechanics\. The baseline executesRollout​\(πθ,x,𝒯\)\\textsc\{Rollout\}\(\\pi\_\{\\theta\},x,\\mathcal\{T\}\)without memory, while the memory\-active path prepends retrieved experiences beforexx\. Because𝒯\\mathcal\{T\}and the calling protocol remain identical across both paths, any performance difference is attributable solely to the experiential guidance rather than altered tool dynamics\.

## 5Experiments

### 5\.1Experimental Setup

#### Benchmarks\.

We evaluate six settings from four benchmarks, covering numerical grounding, multimodal numerical reasoning, temporal\-aware retrieval, and chart understanding\. BizBench SEC\-NUM\(Koncel\-Kedziorskiet al\.,[2023](https://arxiv.org/html/2606.17642#bib.bib11)\)includes 2000 SEC 10\-K/10\-Q items for quantity localization under numeric distractors; we use exact match with 1% relative tolerance\. FinMMR\(Tanget al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib6)\)includes Easy/Medium/Hard splits \(1200/1200/1000\) for bilingual multi\-image financial numerical reasoning; we follow its 0\.2% relative tolerance criterion\. FinTMMBench\(Zhuet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib4)\)includes 1136 temporal\-aware multimodal retrieval\-reasoning items over tables, news, prices, and charts; we report official GPT\-4o\-mini\-judge accuracy\. FinMME\(Luoet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib5)\)includes 2220 chart\-intensive items across diverse financial domains; because the released annotations available to us do not include official per\-sample domain labels required to reproduce the benchmark’s official domain\-normalized average\.Therefore, we supplemented this part ourselves for comparison\. We follow each benchmark’s official split and evaluation protocol, with the FinMME scoring caveat above\. Full dataset, scoring, and implementation details are provided in Appendix[F](https://arxiv.org/html/2606.17642#A6)\. We follow each benchmark’s official split and evaluation protocol\. Full dataset specifications and example items are provided in Appendix[F](https://arxiv.org/html/2606.17642#A6)\.

#### Implementation details\.

We use frozen Qwen3\-VL\-8B\-Instruct\(Baiet al\.,[2025](https://arxiv.org/html/2606.17642#bib.bib10)\)\(temperature 0, DashScope\), serving without augmentation as theBasemodel\-only baseline \(Table[1](https://arxiv.org/html/2606.17642#S5.T1)\)\. As detailed in Section[4\.4](https://arxiv.org/html/2606.17642#S4.SS4), the tool suite \(Base\+FT\) equips the backbone with a ReAct loop \(max 16 steps\) over four tool components for numerical reasoning, data retrieval, visual decoding, and answer gating; its standalone contribution is isolated in the ablation study \(Section[5\.3](https://arxiv.org/html/2606.17642#S5.SS3)\)\. TheFMbank pools∼\\sim2,400 experience accumulation trajectories across all benchmarks, sourced exclusively from training/validation splits held out from test data, and is frozen read\-only at evaluation\. All ablation conditions share identical tool and decoding configurations, isolating the effect ofFMandFT\. Published baseline numbers are collected from respective benchmark sources; exact model variant specifications and per\-cell attributions appear in Appendix[F\.3](https://arxiv.org/html/2606.17642#A6.SS3), with further implementation details in Appendix[F](https://arxiv.org/html/2606.17642#A6)\.

### 5\.2Main Results

Table[1](https://arxiv.org/html/2606.17642#S5.T1)reports per\-benchmark accuracy across all baselines andFinAcumen\. We treat this as a*deployment\-setting comparison*:FinAcumenis a memory\- and tool\-enhanced small\-model system, while published closed\-source baselines are benchmark\-reported native model results without ourFMandFTwrapper\. The claim is not that an 8B backbone is intrinsically stronger than GPT\-4o or Claude, but that external memory and deterministic tools can substantially strengthen a frozen small VLM under constrained deployment\.

The intended reading is vertical: the gap from Base toFinAcumenisolates the framework contribution under the same backbone, while the gap to the best General LLM indicates remaining headroom under the published benchmark setting\. Model variant specifics are documented in Appendix[F\.3](https://arxiv.org/html/2606.17642#A6.SS3)\.

Table 1:Per\-benchmark accuracy across six financial QA settings\. Published baseline entries are benchmark\-reported model results without ourFMandFTwrapper; per\-cell variant details and source attributions are in Appendix[F\.3](https://arxiv.org/html/2606.17642#A6.SS3)\. Best result per column in bold, second\-best underlined\. A dash indicates unreported by the source benchmark\.FinAcumenresults use a frozen Qwen3\-VL\-8B\-Instruct backbone withFMandFT\.Table[1](https://arxiv.org/html/2606.17642#S5.T1)supports three observations\.First,FinAcumenproduces large system\-level gains: the same frozen backbone rises from 27\.67 to 68\.65 on BizBench and from 27\.84 to 51\.08 raw average on FinMME, while also surpassing the published GPT\-4o entries on FinMMR Easy/Medium\.Second, it is more robust than static financial specialization: across all shared columns,FinAcumenoutperforms Llama\-Open\-Finance\-8B, FinLLaVA, and Fin\-R1, suggesting that financial multimodal QA benefits from test\-time evidence routing, verification, and strategy adaptation\.Third, the remaining gaps are informative:FinAcumenstill trails the best General LLM on BizBench, FinMMR Hard, and FinTMMBench, where stronger planning, temporal retrieval, or difficult numerical reasoning is required\. Thus,FinAcumenis best viewed as a deployment\-efficient way to strengthen frozen VLMs, not as evidence that an 8B backbone has superior intrinsic parametric ability\.

### 5\.3Ablation Study

Since Table[1](https://arxiv.org/html/2606.17642#S5.T1)compares a full system against published model\-only baselines, we isolate the source of the gain by togglingFTandFMon the same Qwen3\-VL\-8B backbone\. Table[2](https://arxiv.org/html/2606.17642#S5.T2)reports FinMMR Easy, whose moderate base accuracy \(59\.17\) makes module\-level gains easy to measure\. All variants share the same backbone, decoding configuration, and scoring protocol\.

Table 2:Ablation ofFTandFMon FinMMR Easy with Qwen3\-VL\-8B\-Instruct\.Δ\\Deltadenotes absolute improvement over the Base only variant\.The ablation explains the main\-table gains\. AddingFTraises accuracy from 59\.17 to 74\.08, showing that tools resolve many execution\-level errors\. AddingFMon top of the same tool suite further improves accuracy to 81\.67, a 7\.59\-point gain over tool\-only inference\. This shows thatFinAcumenis not merely a tool wrapper: retrieved experiences provide strategy\-level guidance about tool use, quantity verification, and recurring financial reasoning traps\. Together,FTandFMyield a 22\.50\-point improvement over the base model\.

### 5\.4Memory Accumulation Behavior

To isolate the marginal contribution of experience memory, we evaluate a frozenFMbank on BizBench SEC\-NUM across seven snapshots spanning 0 to 2400 accumulated entries\. Then=0n=0condition corresponds to theFT\-augmented agent with an emptyFMbank and is distinct from the bare model baseline\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x3.png)Figure 3:Accuracy on BizBench SEC\-NUM as a function of accumulatedFMbank size\. Then=0n=0condition includes theFTtool suite\. Per\-step delta annotations reveal diminishing marginal returns, with accuracy approaching a saturation regime beyondn=1800n=1800, marked by the shaded region\.Figure[3](https://arxiv.org/html/2606.17642#S5.F3)reports the accuracy trajectory\. Starting from 55\.15% with no experience entries, accuracy rises to 68\.65% at 2400 entries, a net gain of 13\.5 points\. The curve exhibits diminishing returns: per\-interval gains are most pronounced at smaller bank sizes and progressively decline, entering a saturation regime beyond 1800 entries where additional entries produce marginal improvements\. A detailed decomposition of per\-interval marginal gains is provided in the Appendix\. This monotonic saturating pattern confirms that theFMbank accumulates useful experience during training while selective retrieval at test time prevents stale entries from degrading performance, consistent with the retrieval analysis in Section[5\.5](https://arxiv.org/html/2606.17642#S5.SS5)\.

### 5\.5Sensitivity Analysis

TheFMretrieval pipeline has two hyperparameters: a similarity thresholdτ\\taugoverning entry admission and a retrieval depthkmaxk\_\{\\text\{max\}\}governing how many admitted entries enter the reasoning context\. We analyze each independently\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x4.png)\(a\)Hit rate as a function of bank size, stratified by similarity thresholdτ\\tau\.
![Refer to caption](https://arxiv.org/html/2606.17642v1/x5.png)\(b\)Accuracy as a function of retrieval depthkmaxk\_\{\\text\{max\}\}\.

Figure 4:Sensitivity analysis of two retrieval hyperparameters\. \(a\) Hit rate across bank sizes for different cosine thresholds; three operational regimes emerge, with the practical range centered at moderate filtering strictness\. \(b\) Accuracy as a function of retrieval depth, following an inverted\-U pattern that peaks at moderate depth and degrades at both extremes\. Shaded regions denote 95% bootstrap confidence intervals\.#### Cosine similarity thresholdτ\\tau\.

We evaluateτ\\tauvia hit rate rather than downstream accuracy, as this threshold must establish a viable candidate pool beforekmaxk\_\{\\text\{max\}\}can operate; hit rate measures the fraction of test queries for which at least one bank entry exceeds the threshold\. Figure[4\(a\)](https://arxiv.org/html/2606.17642#S5.F4.sf1)shows hit rates for nineτ\\tauvalues spanning\[0\.50,0\.90\]\[0\.50,0\.90\]against accumulated bank size\. Three operational regimes emerge\. A loose\-filtering regime atτ≤0\.55\\tau\\leq 0\.55exceeds 87% hit rate at 120 entries with negligible semantic filtering\. A strict\-blocking regime atτ≥0\.75\\tau\\geq 0\.75stays below 37% even at 1200 entries, denying most queries access to memory\. A moderate\-filtering regime in\[0\.60,0\.70\]\[0\.60,0\.70\]yields 43% to 80% across bank expansion\. Atτ=0\.65\\tau=0\.65, the center of this moderate regime, hit rate rises from 43\.2% at 120 entries to 79\.8% at 1200 entries; saturation beyond 600 entries mirrors the accumulation plateau observed in Figure[3](https://arxiv.org/html/2606.17642#S5.F3), consistent with early entries capturing prevalent error patterns\.

#### Retrieval depthkmaxk\_\{\\text\{max\}\}\.

We testkmax∈\{1,3,5,7\}k\_\{\\text\{max\}\}\\in\\\{1,3,5,7\\\}atτ=0\.65\\tau=0\.65on 500 questions evenly split across BizBench and FinMMR\-M, with all other configurations matching the main experiments\. Figure[4\(b\)](https://arxiv.org/html/2606.17642#S5.F4.sf2)shows an inverted\-U relationship between retrieval depth and accuracy\. Atkmax=1k\_\{\\text\{max\}\}=1macro accuracy is 57\.80%; expanding to33reaches 57\.85%, a shift of 0\.05 percentage points that lies within bootstrap confidence intervals and is thus not statistically meaningful\. Atkmax=5k\_\{\\text\{max\}\}=5macro accuracy rises to 63\.11%, a gain of 5\.26 points that falls well outside the confidence bounds, with consistent improvements on BizBench of 5\.98 points and FinMMR\-M of 4\.55 points\. This sharp transition is consistent with a scenario where entries ranked fourth and fifth cover strategy categories absent from the top three; the highest\-ranked entries tend to cluster around a single dominant error type, while moderately\-ranked entries span complementary failure modes demanding qualitatively different solution patterns\. Atkmax=7k\_\{\\text\{max\}\}=7macro accuracy drops to 62\.17%, a decline of 0\.94 points, indicating that retrieval beyond five entries introduces distractors that outweigh any additional strategic signal\. The tension between insufficient strategic coverage at lowkmaxk\_\{\\text\{max\}\}and signal dilution at highkmaxk\_\{\\text\{max\}\}is resolved atkmax=5k\_\{\\text\{max\}\}=5, which is adopted for all experiments\.

## 6Conclusion

We introducedFinAcumen, a framework coupling self\-evolving Financial Memory with deterministic Financial Tools for financial multimodal QA\. Across four benchmarks,FinAcumenwith a frozen 8B VLM achieves consistent gains over finance\-specialized models and competitive performance against leading general\-purpose systems at a fraction of their parameter scale\. Ablation confirms complementary roles:FMimproves routing and strategy selection via polarity\-aware memory activation, whileFTguarantees numerical fidelity through execution\-environment grounding; when no relevant experience is retrievable,FMfalls back to the tool\-augmented baseline without degradation\. The decoupled architecture, in which an evolving experience bank operates orthogonally to a deterministic tool suite around a frozen base model, suggests a general paradigm for professional\-domain AI: externalized memory captures reusable strategy patterns without destabilizing the core model, while domain\-specific tools provide computational guarantees that generative components alone typically struggle to guarantee\. Future work spans two dimensions: sample\-efficient consolidation and compositional memory structures to reduce cold\-start cost and enable cross\-task strategy transfer; and cross\-lingual alignment with stronger visual encoders for robust deployment on multilingual and chart\-dense real\-world settings\.

## References

- Self\-rag: learning to retrieve, generate, and critique through self\-reflection\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen,et al\.\(2025\)Qwen3\-VL technical report\.arXiv preprint arXiv:2511\.21631\.Note:Technical report for Qwen3\-VL; base model is Qwen3\-VL\-8B\-Instruct\.Cited by:[§5\.1](https://arxiv.org/html/2606.17642#S5.SS1.SSS0.Px2.p1.1)\.
- G\. Caillaut, R\. Qader, J\. Liu, M\. Nakhlé, A\. Sadoune, M\. Ahmim, and J\. Barthelemy \(2025\)The llm pro finance suite: multilingual large language models for financial applications\.arXiv preprint arXiv:2511\.08621\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p2.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Dai, Y\. Lin, Z\. Xie, and Y\. Wang \(2026\)RealFin: how well do llms reason about finance when users leave things unsaid?\.arXiv preprint arXiv:2602\.07096\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p2.1)\.
- S\. Deng, Z\. Wang, R\. Mao, C\. D\. Giurcăneanu, and J\. Liu \(2026\)CLER: improving multimodal financial reasoning by cross\-mllm error reflection\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 30494–30502\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p2.1)\.
- J\. Huang, M\. Xiao, D\. Li, Z\. Jiang, Y\. Yang, Y\. Zhang, L\. Qian, Y\. Wang, X\. Peng, Y\. Ren,et al\.\(2024\)Open\-finllms: open multimodal large language models for financial applications\.arXiv preprint arXiv:2408\.11878\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p2.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1)\.
- R\. Koncel\-Kedziorski, M\. Krumdick, V\. Lai, V\. Reddy, C\. Lovering, and C\. Tanner \(2023\)BizBench: a quantitative reasoning benchmark for business and finance\.arXiv preprint arXiv:2311\.06602\.Note:SEC\-grounded numerical reasoning benchmark; cited in §4\.1 dataset table\.Cited by:[§F\.3](https://arxiv.org/html/2606.17642#A6.SS3.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p1.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2606.17642#S5.SS1.SSS0.Px1.p1.1)\.
- N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang \(2024\)Lost in the middle: how language models use long contexts\.Transactions of the association for computational linguistics12,pp\. 157–173\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Liu, X\. Guo, Z\. Yang, F\. Lou, L\. Zeng, J\. Niu, M\. Li, Q\. Qi, Z\. Liu, Y\. Han,et al\.\(2025\)Fin\-r1: a large language model for financial reasoning through reinforcement learning\.arXiv preprint arXiv:2503\.16252\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p2.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Luo, Z\. Kou, L\. Yang, X\. Luo, J\. Huang, Z\. Xiao, J\. Peng, C\. Liu, J\. Ji, X\. Liu,et al\.\(2025\)Finmme: benchmark dataset for financial multi\-modal reasoning evaluation\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 29465–29489\.Cited by:[§F\.3](https://arxiv.org/html/2606.17642#A6.SS3.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p2.1),[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2606.17642#S5.SS1.SSS0.Px1.p1.1)\.
- S\. Ouyang, J\. Yan, I\. Hsu, Y\. Chen, K\. Jiang, Z\. Wang, R\. Han, L\. T\. Le, S\. Daruki, X\. Tang,et al\.\(2025\)Reasoningbank: scaling agent self\-evolving with reasoning memory\.arXiv preprint arXiv:2509\.25140\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Shi, X\. Chen, K\. Misra, N\. Scales, D\. Dohan, E\. H\. Chi, N\. Schärli, and D\. Zhou \(2023\)Large language models can be easily distracted by irrelevant context\.InInternational Conference on Machine Learning,pp\. 31210–31227\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.Advances in neural information processing systems36,pp\. 8634–8652\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Tang, E\. Haihong, R\. Li, J\. Liu, L\. Jia, Z\. Hao, Z\. Yang, Y\. Li, H\. Tian, X\. Hu,et al\.\(2026\)FinMMDocR: benchmarking financial multimodal reasoning with scenario awareness, document understanding, and multi\-step computation\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 25858–25866\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p2.1)\.
- Z\. Tang, J\. Liu, Z\. Yang, R\. Li, Z\. Rong, H\. He, Z\. Hao, X\. Hu, K\. Ji, Z\. Ma,et al\.\(2025\)Finmmr: make financial numerical reasoning more multimodal, comprehensive, and challenging\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 3245–3257\.Cited by:[§F\.3](https://arxiv.org/html/2606.17642#A6.SS3.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p2.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2606.17642#S5.SS1.SSS0.Px1.p1.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao \(2023\)ReAct: synergizing reasoning and acting in language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- C\. Zhang, J\. He, J\. He, K\. Sycara, and Y\. Xie \(2026\)Evolving contextual safety in multi\-modal large language models via inference\-time self\-reflective memory\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,Note:EchoSafe: free\-form NL insight memory with cosine\-similarity retrieval and simple\-append \(no deduplication or pruning\); cited in ablation as the−\-dedup/prune baseline\. arXiv:2603\.15800\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. Huang \(2024\)Expel: llm agents are experiential learners\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 19632–19642\.Cited by:[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Zhu, J\. Li, L\. Pan, W\. Wang, F\. Feng, C\. Wang, H\. Luan, and T\. Chua \(2025\)Towards temporal\-aware multi\-modal retrieval augemented generation in finance\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 6289–6297\.Cited by:[§F\.3](https://arxiv.org/html/2606.17642#A6.SS3.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p1.1),[§1](https://arxiv.org/html/2606.17642#S1.p2.1),[§1](https://arxiv.org/html/2606.17642#S1.p3.1),[§2](https://arxiv.org/html/2606.17642#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2606.17642#S5.SS1.SSS0.Px1.p1.1)\.

Appendix

A Notation\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[A](https://arxiv.org/html/2606.17642#A1)

B Inference Algorithm\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B](https://arxiv.org/html/2606.17642#A2)

C Marginal Gain Decomposition of Memory Accumulation\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[C](https://arxiv.org/html/2606.17642#A3)

D Memory Strategy Ablation\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[D](https://arxiv.org/html/2606.17642#A4)

E Retrieval Stability Analysis\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E](https://arxiv.org/html/2606.17642#A5)

E\.1 Experimental Design\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E\.1](https://arxiv.org/html/2606.17642#A5.SS1)

E\.2 Retrieval Metrics\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E\.2](https://arxiv.org/html/2606.17642#A5.SS2)

E\.3 Coverage and Retrieval Density\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E\.3](https://arxiv.org/html/2606.17642#A5.SS3)

E\.4 Retrieval Stability\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E\.4](https://arxiv.org/html/2606.17642#A5.SS4)

F Experimental Details\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F](https://arxiv.org/html/2606.17642#A6)

F\.1 Evaluation Protocols\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F\.1](https://arxiv.org/html/2606.17642#A6.SS1)

F\.2 FinAcumen Configuration\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F\.2](https://arxiv.org/html/2606.17642#A6.SS2)

F\.3 Baseline Attribution\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F\.3](https://arxiv.org/html/2606.17642#A6.SS3)

G Experimental Details\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[G](https://arxiv.org/html/2606.17642#A7)

H Limitations\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[H](https://arxiv.org/html/2606.17642#A8)

## Appendix ANotation

We define mathematical symbols that appear across multiple sections of the main text\. Symbols introduced and used only within a single paragraph or algorithm are defined locally and not repeated here\.

Table 3:Global mathematical notation\.
## Appendix BInference Algorithm

Algorithm 1FinAcumensingle\-instance inference with retrieval1:

xx,

ℳ\\mathcal\{M\},

πθ\\pi\_\{\\theta\},

𝒯\\mathcal\{T\},

ω\\omega,

EmgrE\_\{\\mathrm\{mgr\}\}
2:

vq←E​\(q​\(x\)⊕c​\(x\)\[:600\]\)v\_\{q\}\\leftarrow E\(q\(x\)\\oplus c\(x\)\_\{\[:600\]\}\)
3:

Vℳ←Emgr\.resolve\_all​\(ℳ\)V\_\{\\mathcal\{M\}\}\\leftarrow E\_\{\\mathrm\{mgr\}\}\.\\text\{resolve\\\_all\}\(\\mathcal\{M\}\)
4:

s→←vq⋅Vℳ⊤\\vec\{s\}\\leftarrow v\_\{q\}\\cdot V\_\{\\mathcal\{M\}\}^\{\\top\}
5:

I←\{i∣si≥0\.65\}I\\leftarrow\\\{i\\mid s\_\{i\}\\geq 0\.65\\\}
6:if

I=∅I=\\emptysetthen

7:return

Rollout​\(πθ,x,𝒯\)\\textsc\{Rollout\}\(\\pi\_\{\\theta\},x,\\mathcal\{T\}\)
8:endif

9:

I→←DescendingSort​\(I,s→\)\\vec\{I\}\\leftarrow\\textsc\{DescendingSort\}\(I,\\vec\{s\}\)
10:

I→←DedupBySource​\(I→\)\\vec\{I\}\\leftarrow\\textsc\{DedupBySource\}\(\\vec\{I\}\)
11:

I→top←I→\[:3\]\\vec\{I\}\_\{\\text\{top\}\}\\leftarrow\\vec\{I\}\_\{\[:3\]\}
12:

xml←RenderMemoryBlock​\(ℳ​\[I→top\]\)\\text\{xml\}\\leftarrow\\textsc\{RenderMemoryBlock\}\(\\mathcal\{M\}\[\\vec\{I\}\_\{\\text\{top\}\}\]\)
13:

y←Rollout​\(πθ,Prepend​\(xml,x\),𝒯\)y\\leftarrow\\textsc\{Rollout\}\\\!\\big\(\\pi\_\{\\theta\},\\textsc\{Prepend\}\(\\text\{xml\},x\),\\mathcal\{T\}\\big\)
14:

UpdateStats​\(I→top,y\)\\textsc\{UpdateStats\}\(\\vec\{I\}\_\{\\text\{top\}\},y\)
15:if

ω\\omegathen

16:schedule async

Consolidate​\(x,y,ℳ\)\\textsc\{Consolidate\}\(x,y,\\mathcal\{M\}\)
17:endif

18:return

yy

## Appendix CMarginal Gain Decomposition of Memory Accumulation

Figure[5](https://arxiv.org/html/2606.17642#A3.F5)decomposes the memory accumulation curve from Section[5\.4](https://arxiv.org/html/2606.17642#S5.SS4)into per\-segment marginal gains, defined as the accuracy improvement per 100 accumulatedFMentries for each consecutive snapshot interval\. The evaluation protocol mirrors the main analysis: theFT\-augmented Qwen3\-VL\-8B\-Instruct backbone is evaluated on the full 2000\-item BizBench SEC\-NUM test set at seven bank snapshots spanning 0 to 2400 entries\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x6.png)Figure 5:Marginal accuracy gain per 100 accumulatedFMentries on BizBench SEC\-NUM\. The curve traces a diminishing\-returns pattern across consecutive snapshot intervals\. The dashed vertical line separates the main accumulation phase from the saturation regime\.The marginal gain decomposition in Figure[5](https://arxiv.org/html/2606.17642#A3.F5)reveals a three\-phase accumulation dynamic on BizBench\.Phase 1 \(0–120 entries\):The initial accumulation of 120 entries yields a marginal gain of \+1\.50 per 100 entries, capturing the most common error patterns\.Phase 2 \(240–1800 entries\):The next 1560 entries sustain a robust marginal efficiency of \+0\.51 to \+1\.11 per 100 entries, with the 240–600 interval producing the strongest per\-entry return \(\+1\.11\) and the 600–1200 interval yielding \+0\.51\. The 1200–1800 interval yields \+0\.57 per 100 entries, reflecting an uptick in marginal efficiency as newly accumulated entries target previously uncovered error patterns\.Phase 3 \(n≥1800n\\geq 1800\):Beyond 1800 entries, the marginal gain declines to \+0\.22 per 100 entries, or approximately \+1\.30 points across the final 600\-entry block\. This diminishing return supports the selection of 1800 entries as the saturation knee for the production bank\.

The consistently positive marginal gains across all intervals, including the saturation phase, confirm that theFMbank continues to provide useful signal even after substantial accumulation\. While individual snapshots contain minor fluctuations \(the 120–240 interval shows a near\-zero change of−0\.04\-0\.04\), the overall trajectory is monotonically increasing on the macro timescale\. This characteristic supports our deployment strategy of accumulating experience passively throughout training and freezing the bank at the empirically determined saturation point for test\-time inference, as described in Section[5\.4](https://arxiv.org/html/2606.17642#S5.SS4)\.

## Appendix DMemory Strategy Ablation

This appendix isolates the marginal contribution of each component in theFinAcumenpipeline by comparing five injection strategies that toggle four fields of the retrieved memory block\.

- •Q\(Question\): the original question text of the top\-retrieved similar item\.
- •A\(Answer\): the gold answer of the retrieved item\.
- •E\(Experience\): distilled Findings, reusable solution recipes, and Cautions, guard rules from past failures\.
- •Annot\(Annotation\): LLM\-generated personalized directives that map generic rules to query\-specific entities such as ticker symbols, date ranges, and metric names\.

All strategies share Qwen3\-VL\-8B\-Instruct as the backbone and operate on the same 600\-question test subsample with MIN\_COSINE set to 0\.60\. Strategies A and B use the single\-shot baseline\-raw variant without tools; strategies C through E use the full FinAcumen variant with the four\-tool ReAct loop\. Table[4](https://arxiv.org/html/2606.17642#A4.T4)reports per\-benchmark accuracy and the macro average across the four benchmarks\.

Table 4:Memory strategy ablation\. Checkmarks indicate which fields are injected into the memory block for each strategy\. Macro is the arithmetic mean across the four benchmarks\.The strategy comparisons in Table[4](https://arxiv.org/html/2606.17642#A4.T4)decompose the contribution of each field\. Adding gold answers to Q\-only references, A to B, yields 3\.25 macro points, confirming that reference answers carry a measurable signal\. Adding tools and annotation without experience, B to D, yields 2\.25 points, less than half the gain from experience alone\.

Experience provides the dominant marginal improvement\. Strategy C, which includes Q, A, and experience but omits annotation, exceeds strategy D, which includes Q, A, and annotation but omits experience, by 3\.50 macro points\. This gap isolates Findings and Cautions from the effect of annotation\. The gain is concentrated on FinMME, from 42 to 47, and on FinMMR Easy, from 64 to 71, benchmarks on which sign convention and unit conversion errors recur systematically and are captured by distilled guard rules\.

Annotation amplifies experience rather than adding independently to it\. Strategy E, the full pipeline, exceeds C by 5\.50 points, a margin larger than the D\-to\-C gap of 3\.50\. This asymmetry indicates that annotator directives are most effective when experience content is already rich: personalization prevents a guard rule about fiscal year 2022 from being misapplied to a query about fiscal year 2024, but generic Findings and Cautions must first be present to be personalized\. The total gain from A to E is 14\.50 macro points, from 53\.00 to 67\.50\. The monotonic progression across all five strategies confirms that tools, experience, and annotation address complementary failure modes that no subset of the pipeline can cover alone\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x7.png)Figure 6:Memory strategy ablation on a shared test subsample\. Checkmarks in Table[4](https://arxiv.org/html/2606.17642#A4.T4)indicate the injected fields for each strategy\.
## Appendix ERetrieval Stability Analysis

### E\.1Experimental Design

We evaluate retrieval composition across 11 bank snapshots, from 0 to 1200 entries, by computing cosine similarity between all 11,379 test query embeddings and each bank entry at the operating thresholdτ=0\.65\\tau=0\.65\. The snapshots are constructed as cumulative prefixes drawn uniformly across six benchmarks: entries fromfinacumen/memory/v8\_unified/meta\.jsonare partitioned bysource\.dataset, shuffled with seed 42, and the firstS/6S/6entries per benchmark are concatenated for each target sizeS∈\{0,120,240,…,1200\}S\\in\\\{0,120,240,\\dots,1200\\\}\. Each successive snapshot retains all entries from the previous size and appends 20 additional entries per benchmark, yielding a cumulative prefix of 120 entries per step\. A small fraction of entries lack resolvable embeddings, yielding actual bank cardinalities slightly below the nominal target sizes\. While we fix the primary operational threshold atτ=0\.65\\tau=0\.65for cross\-benchmark analysis, we also sweepτ∈\[0\.50,0\.90\]\\tau\\in\[0\.50,0\.90\]to examine how retrieval density and volatility scale under different matching strictness\.

### E\.2Retrieval Metrics

We report four metrics, whereRn​\(q\)R\_\{n\}\(q\)denotes the set of entries retrieved for queryqqat snapshotnnandQQdenotes the full query set of size 11,379\.

Hn=1\|Q\|​∑q∈Q𝟏​\[\|Rn​\(q\)\|\>0\]H\_\{n\}=\\frac\{1\}\{\|Q\|\}\\sum\_\{q\\in Q\}\\mathbf\{1\}\[\|R\_\{n\}\(q\)\|\>0\]
Hit rateHnH\_\{n\}measures the fraction of queries receiving at least one retrieved entry at bank sizenn\.

E¯n=1\|Q\|​∑q∈Q\|Rn​\(q\)\|\\bar\{E\}\_\{n\}=\\frac\{1\}\{\|Q\|\}\\sum\_\{q\\in Q\}\|R\_\{n\}\(q\)\|
Mean entries per queryE¯n\\bar\{E\}\_\{n\}captures retrieval density\.

For adjacent snapshotsn1n\_\{1\}andn2n\_\{2\}, the per\-query Jaccard similarity is defined with an explicit case for two empty retrieval sets:

Jq​\(n1,n2\)=\{1,if​Rn1​\(q\)=Rn2​\(q\)=∅,\|Rn1​\(q\)∩Rn2​\(q\)\|\|Rn1​\(q\)∪Rn2​\(q\)\|,otherwise\.J\_\{q\}\(n\_\{1\},n\_\{2\}\)=\\left\\\{\\begin\{array\}\[\]\{l@\{\}l\}1,\\quad\\kern 5\.0pt\\kern 5\.0pt\\text\{if \}R\_\{n\_\{1\}\}\(q\)=R\_\{n\_\{2\}\}\(q\)=\\emptyset,\\\\ \\dfrac\{\|R\_\{n\_\{1\}\}\(q\)\\cap R\_\{n\_\{2\}\}\(q\)\|\}\{\|R\_\{n\_\{1\}\}\(q\)\\cup R\_\{n\_\{2\}\}\(q\)\|\},\\kern 5\.0pt\\;\\text\{otherwise\}\.\\end\{array\}\\right\.
Jn1,n2=1\|Q\|​∑q∈QJq​\(n1,n2\)J\_\{n\_\{1\},n\_\{2\}\}=\\frac\{1\}\{\|Q\|\}\\sum\_\{q\\in Q\}J\_\{q\}\(n\_\{1\},n\_\{2\}\)
Adjacent\-size Jaccard similarityJn1,n2J\_\{n\_\{1\},n\_\{2\}\}quantifies compositional stability across snapshots\. AssigningJq=1J\_\{q\}=1when both retrieval sets are empty inflates Jaccard estimates for benchmarks with low hit rates, a phenomenon of trivial stability discussed in Section[E\.4](https://arxiv.org/html/2606.17642#A5.SS4)\.

Cn1,n2=1\|Q\|​∑q∈Q𝟏​\[Rn1​\(q\)≠Rn2​\(q\)\]C\_\{n\_\{1\},n\_\{2\}\}=\\frac\{1\}\{\|Q\|\}\\sum\_\{q\\in Q\}\\mathbf\{1\}\[R\_\{n\_\{1\}\}\(q\)\\neq R\_\{n\_\{2\}\}\(q\)\]
Change fractionCn1,n2C\_\{n\_\{1\},n\_\{2\}\}captures retrieval volatility as the strict proportion of queries whose retrieval set differs between snapshots\.

Per\-dataset hit rate and Jaccard curves, together with their global aggregates, appear in Figures[7](https://arxiv.org/html/2606.17642#A5.F7)and[9](https://arxiv.org/html/2606.17642#A5.F9)\. Table[5](https://arxiv.org/html/2606.17642#A5.T5)summarizes the terminal\-state metrics at bank size 1200 forτ=0\.65\\tau=0\.65\.

Table 5:Terminal retrieval state at bank sizeN=1200N=1200underτ=0\.65\\tau=0\.65\.H1200H\_\{1200\}denotes hit rate,E¯1200\\bar\{E\}\_\{1200\}mean entries per query, andJ1080,1200J\_\{1080,1200\}the Jaccard similarity at the final adjacent\-size transition\.![Refer to caption](https://arxiv.org/html/2606.17642v1/x8.png)Figure 7:Hit rate as a function ofFMbank size atτ=0\.65\\tau=0\.65, with a global aggregate curve\. All benchmarks exhibit saturation beyond 600 entries; cross\-dataset variation reflects divergent semantic affinity between benchmark query distributions and the bank’s memorized patterns\.
### E\.3Coverage and Retrieval Density

Hit rate saturates beyond 600 entries across all benchmarks: FinTMM reaches 90\.4% at 1200 entries whereas FinMME reaches only 47\.5%\. This gap reflects divergent semantic affinity between benchmark query distributions and the patterns memorized in the bank\. Chart\-oriented FinMME queries activate substantially fewer stored experiences than numerical reasoning queries, consistent with a bank composition that over\-represents text\- and table\-centric failure cases\. Mean entries per query confirms this pattern: atτ=0\.65\\tau=0\.65, FinMMR\-Medium and FinMMR\-Hard retrieve 54\.5 and 38\.2 entries per query on average, while FinMME retrieves 1\.1\. FinTMM presents an intermediate profile with a hit rate of 90\.4% but a mean entry count of only 4\.0, indicating that its queries resolve with sparser yet highly targeted retrieval sets\.

Under threshold\-based retrieval, the expected number of retrieved entries followsE​\[\|Rn​\(q\)\|\]=n⋅pE\[\|R\_\{n\}\(q\)\|\]=n\\cdot p, wherep=P​\(sim​\(q,e\)≥τ\)p=P\(\\text\{sim\}\(q,e\)\\geq\\tau\)is independent of the current bank size\. This produces the nearly linear growth of mean entries observed in Figure[8](https://arxiv.org/html/2606.17642#A5.F8)\. Because new entries are purely additive under threshold filtering, the retrieval set monotonically expands:Rn​\(q\)⊆Rn\+Δ​n​\(q\)R\_\{n\}\(q\)\\subseteq R\_\{n\+\\Delta n\}\(q\)holds for all queries\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x9.png)Figure 8:Mean number of retrieved entries per query aggregated across all 11,379 test queries\. Each curve corresponds to one cosine threshold; the near\-linear growth is predicted byE​\[\|Rn​\(q\)\|\]=n⋅pE\[\|R\_\{n\}\(q\)\|\]=n\\cdot p\.
### E\.4Retrieval Stability

SinceRn​\(q\)⊆Rn\+Δ​n​\(q\)R\_\{n\}\(q\)\\subseteq R\_\{n\+\\Delta n\}\(q\)holds under threshold retrieval, the expected Jaccard similarity between adjacent snapshots has the closed form

E​\[Jn,n\+Δ​n\]≈n⋅p\(n\+Δ​n\)⋅p=nn\+Δ​n,E\[J\_\{n,n\+\\Delta n\}\]\\approx\\frac\{n\\cdot p\}\{\(n\+\\Delta n\)\\cdot p\}=\\frac\{n\}\{n\+\\Delta n\},which increases monotonically and approaches 1\.0 asnngrows\. This matches the dominant trend in Figure[9](https://arxiv.org/html/2606.17642#A5.F9): all benchmarks exhibit a monotonic rise toward saturation, with the final transition values exceeding 0\.90 for every dataset\. Minor terminal fluctuations, such as the 0\.004 drop observed in FinMMR\-Medium, are consistent with sample variance: at specific transitions the empirical retrieval probabilityp^\\hat\{p\}of newly admitted entries deviates slightly from the populationpp, producing sub\-thousandth\-level oscillations that are not statistically significant\.

The high Jaccard values of FinMME require careful interpretation\. FinMME exhibits the lowest hit rate among all benchmarks, reaching only 47\.5% at 1200 entries\. For the remaining 52\.5% of FinMME queries no entries are retrieved at either snapshot, and the conventionJq​\(∅,∅\)=1J\_\{q\}\(\\emptyset,\\emptyset\)=1applies\. This trivial stability inflates the observed Jaccard toward 1\.0 and partially accounts for FinMME’s higher Jaccard curve\. For benchmarks with sufficient retrieval coverage the Jaccard trajectory reflects genuine compositional convergence\.

The change fractionCn1,n2C\_\{n\_\{1\},n\_\{2\}\}measures the proportion of queries for which at least one newly admitted entry exceeds the similarity threshold\. Under threshold retrieval, a query’s retrieval set changes if and only if at least one entry among theΔ​n=120\\Delta n=120newly added entries satisfies the similarity condition\. This probability is

P​\(Change\)=1−\(1−p\)Δ​n,P\(\\text\{Change\}\)=1\-\(1\-p\)^\{\\Delta n\},which depends solely onτ\\tauand the fixed step sizeΔ​n\\Delta n, not on the current bank sizenn\. Empirically, the change fraction curves in Figure[10](https://arxiv.org/html/2606.17642#A5.F10)are near\-horizontal across all bank size transitions for every threshold value, confirming this invariance\.

![Refer to caption](https://arxiv.org/html/2606.17642v1/x10.png)Figure 9:Adjacent\-bank Jaccard similarity atτ=0\.65\\tau=0\.65with error bars indicating one standard deviation across queries\. Jaccard similarity increases monotonically toward 1\.0, consistent withE​\[J\]≈n/\(n\+Δ​n\)E\[J\]\\approx n/\(n\+\\Delta n\); minor local fluctuations are attributable to sample variance in newly added entries\.![Refer to caption](https://arxiv.org/html/2606.17642v1/x11.png)Figure 10:Change fraction across adjacent bank snapshots for nine cosine thresholds\. The near\-horizontal trajectory for each threshold confirms that retrieval volatility is invariant to bank size, as predicted byP​\(Change\)=1−\(1−p\)Δ​nP\(\\text\{Change\}\)=1\-\(1\-p\)^\{\\Delta n\}\.Taken together, these results establish that under threshold\-based retrieval, retrieval stability is an inherent consequence of the retrieval mechanism rather than an emergent property of bank size\. Jaccard similarity converges toward 1\.0 asn/\(n\+Δ​n\)n/\(n\+\\Delta n\)approaches unity, and change fraction remains constant at1−\(1−p\)Δ​n1\-\(1\-p\)^\{\\Delta n\}independent ofnn\. For practitioners, adjusting the similarity thresholdτ\\tauoffers a more direct lever for controlling retrieval churn than limiting bank size\.

## Appendix FExperimental Details

This appendix documents the experimental protocol behind Table[1](https://arxiv.org/html/2606.17642#S5.T1)and the ablation analyses\. The goal is to make the comparison reproducible while keeping the main text focused on the central findings\.

### F\.1Evaluation Protocols

We evaluateFinAcumenon the official test split of each benchmark and follow the native metric defined by the corresponding benchmark paper\. For BizBench SEC\-NUM, predictions are normalized by removing currency markers, percentage signs, commas, and whitespace before exact\-match comparison with a 1% relative numerical tolerance\. For FinMMR, we report accuracy on the Easy, Medium, and Hard splits using the official 0\.2% relative tolerance after numeric and unit normalization\. For FinTMMBench, we use the benchmark’s GPT\-4o\-mini judge to decide whether the prediction and the reference answer express the same financial fact\. For FinMME, we report answer accuracy, corresponding to the core correctness component of FinScore: multiple\-choice questions require exact letter\-set agreement and numerical questions follow the tolerance specified by each item\.

All reportedFinAcumenresults are computed as the fraction of correctly answered test items under these native scoring rules\. We do not introduce a unified post\-hoc metric across datasets because the benchmarks differ substantially in answer form, ranging from exact quantities to temporally grounded retrieval answers and chart\-based multiple\-choice questions\.

Table 6:Benchmark settings used in the main experiments\. Each benchmark keeps its original split and native scoring rule\.Table 7:Per\-cell model variants for grouped General LLM baselines in Table[1](https://arxiv.org/html/2606.17642#S5.T1)\. All numbers are copied from the corresponding benchmark papers\. A dash indicates that the source benchmark did not report the model\.
### F\.2FinAcumen Configuration

The backbone for allFinAcumenand Base rows is Qwen3\-VL\-8B\-Instruct\. The backbone parameters are frozen throughout;FinAcumenchanges only the inference\-time context and tool execution path throughFMandFT\. Unless otherwise stated, decoding uses temperature 0 for final test evaluation, and the ReAct\-style solver is capped at 16 tool\-interaction steps\. The tool suite is fixed across allFinAcumenruns and consists of Python execution for deterministic arithmetic, financial data lookup, OCR\-based visual structure extraction, and a termination gate that checks source traceability, unit consistency, and answer format\.

The memory bank is built before final testing from training/validation trajectories and is kept read\-only during evaluation\. Test questions never write new entries into the bank\. Each memory entry stores the source question, gold answer, distilled findings from successful trajectories, and cautions extracted from failed trajectories\. At inference time, the query embedding is formed from the question and the first 600 characters of the context, then compared against pre\-computed memory\-entry embeddings\. Entries below the cosine thresholdτ=0\.65\\tau=0\.65are discarded, duplicate source items are removed, and at mostkmax=5k\_\{\\max\}=5retrieved entries are rendered into the memory prefix for the main results in Table[1](https://arxiv.org/html/2606.17642#S5.T1)\. The Base row disables bothFMandFT, while ablations keep the backbone, decoding configuration, tool interfaces, and benchmark scoring fixed\. All other retrieval and tool parameters follow the architecture described in Section[4](https://arxiv.org/html/2606.17642#S4)\.

### F\.3Baseline Attribution

The non\-FinAcumenbaseline numbers in Table[1](https://arxiv.org/html/2606.17642#S5.T1)are taken from the original benchmark papers rather than re\-run in our environment: BizBench results fromKoncel\-Kedziorskiet al\.\([2023](https://arxiv.org/html/2606.17642#bib.bib11)\), FinMMR results fromTanget al\.\([2025](https://arxiv.org/html/2606.17642#bib.bib6)\), FinTMMBench results fromZhuet al\.\([2025](https://arxiv.org/html/2606.17642#bib.bib4)\), and FinMME results fromLuoet al\.\([2025](https://arxiv.org/html/2606.17642#bib.bib5)\)\. This choice follows the standard leaderboard\-comparison protocol for these benchmarks and avoids conflating our framework contribution with differences in closed\-source API versions, prompting details, or evaluation access\. When a benchmark paper reports a model family rather than a single shared variant across all columns, Table[1](https://arxiv.org/html/2606.17642#S5.T1)groups the results by family and the per\-cell model variant is specified below\.

For Qwen2\.5\-VL\-72B, Gemini 2\.0 Flash, and the financial\-domain baselines, we preserve the exact variants and scores reported by the benchmark sources whenever available\. Missing cells in Table[1](https://arxiv.org/html/2606.17642#S5.T1)mean that the benchmark paper did not report that model under the official setting, not that the model failed evaluation\.

## Appendix GPrompt Templates

![Refer to caption](https://arxiv.org/html/2606.17642v1/x12.png)Figure 11:Prompt orchestration inFinAcumen\.Figure[11](https://arxiv.org/html/2606.17642#A7.F11)provides a visual overview of the prompt orchestration inFinAcumen\. Each prompt excerpt is labeled by its corresponding agent role, described in §[4\.3](https://arxiv.org/html/2606.17642#S4.SS3)\.

Solver agent \(FinAcumen System Prompt, excerpted\)[⬇](data:text/plain;base64,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)<Memory\_Handling\>Theinputmaycontaina<Memory\_Block\>beforethe<Problem\>block\.Each<Entry\>has:\-<Question\>:Theoriginalproblemthatproducedthisexperience\.Useittojudgewhetheryourcurrentproblemisinasimilarscenario\(industry,datasource,problemshape\)\.\-<Answer\>:Thecorrectanswertothatoriginalproblem\.Useasaformatreference\-\-\-nottocopythevalue,buttounderstandwhatkindofoutputyourproblemexpects\(number,text,percentage,etc\.\)\.\-<Experience\>:Distilledreusablerulesorganizedas<Findings\>and<Cautions\>:<Findings\>:Patternsthatsucceeded\-\-\-followthesestrategies\.<Cautions\>:Guardrulesfromfailures\-\-\-applyiftheconditiontriggers\.Theblockendswithan<OptOut\>\-\-\-useitonlywhenanentrytrulydoesnotfit\.</Memory\_Handling\><Think\_Steps\><Think\>\#\#\#1\.InputInventory\(MANDATORY\-\-\-doNOTskiptocomputation\)Outputacompleteinventorybeforeanytoolcall\.\[NARRATIVE\]Everystandalonenumberwithitsunitandcontext,oneperline\.\[TABLES\]ALLrowlabelswithALLcolumnvalues;listeachcell\.\[IMAGES\]ALLfieldsfromOCR/vision;flagambiguousreadings\.\[TERM\_MATCH\]Mapquestiontermstoinventoryentries\.IfatermhasNOdirectmatch,listcandidatematchesexplicitly\.Ifanysectionisincomplete,youhavenotfinishedreading\.DONOTproceedtoStrategyuntiltheinventorycoversALLofContext\.\#\#\#2\.ExperienceApplicability\(ifMemory\_Blockpresent\)\{step2\_text\}\#\#\#3\.ProblemUnderstanding\-Whatquantity?Unit?Precision?CheckInstruction\.\-ResolveTERM\_MATCHwarnings\.\-OutputTYPE:ENTITYNAME,NUMERICVALUE,DESCRIPTIVESENTENCE,VERDICT,orNEWSSENTIMENT\.Matchanswerformattoexpectedtype:\*ENTITY\-\>terminate\(final\_answer="CompanyName"\)\*NUMERIC\-\>terminate\(final\_answer="VALUEUNIT"\)\*DESCRIPTIVE\-\>terminate\(final\_answer="Direction:descriptionwithnumbers"\)\*VERDICT\-\>terminate\(final\_answer="Yes/No"or"true/false"\)\*NEWS\-\>terminate\(final\_answer="Yes"or"No"\)\#\#\#4\.StrategyWriteALLvaluesaspythonvariablesbeforecomputing\(e\.g\.,rev\_2022=125\.4\)\.Checkunitconversion\(millions<\>billions,%<\>decimal\)\.\#\#\#5\.NextActionToolcallorterminate\.</Think\>After</Think\>,emitthetoolcall\.No&lt;Answer&gt;block\.</Think\_Steps\><Invariants\>\-NEVERskipnarrativeparagraphstojumptotables\.Contextisonedocument\-\-\-proseANDtablesmustbothberead\.\-"datanotavailable"isnevertrue\.Ifyouthinkdataismissing,re\-readfromthebeginning\-\-\-youmissedasection\.\-NEVERsubmitplaceholdertextasfinal\_answer:avoid’Iwillterminate’,’datanotavailable’,’insufficientdata’,emptystrings\.\-final\_answerMUSTcontaintheactualanswervalue,notameta\-status\.</Invariants\>

Summary agent—Experience synthesis \(excerpted\)[⬇](data:text/plain;base64,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)<findings\_guide\>Writegeneralizedreusablestrategies\.StripALLspecificvalues:numbers,years,entitynames\.Format:"To<goal\>,use/do<method\>\."Trigger:whatinformationformattriggersthis?Whatquestiontype?Method:whattodostepbystep?Whichtool?Whichformula?Eachfindingisonesentence,self\-containedandexecutable\.<cautions\_guide\>Writegeneralizedguardrules\.StripALLspecificvalues\.Format:"When<detectablecondition\>,<correctiveaction\>\."Conditionmustbedetectablefromquestion/contexttext\(nothindsight\)\.Actionmustbeconcrete:whattocheck,howtocheck,whattodo\.<Output\>ReturnONLY:<Think\>analysis:\[problem\-specificderivation\]\-\>Fromabove,generalizefindings:\-\>Fromabove,generalizecautions:\-\>Verify:doeseachitemhavetracesupport?Istheconditiondetectable?</Think\><Answer\>\{"findings":\["To\.\.\.,use/do\.\.\."\],"cautions":\["When\.\.\.,\.\.\."\],"analysis":"<<REPLACE\_WITH\_YOUR\_PROBLEM\_DERIVATION\>\>"\}</Answer\>

Annotator agent—Directive annotation \(excerpted\)[⬇](data:text/plain;base64,PFRhc2s+CkZvciBlYWNoIGVudHJ5IHdyaXRlIDEtMyBkaXJlY3RpdmUgc2VudGVuY2VzIHVzaW5nIE1VU1QgLyBETyBOT1QuCkV4dHJhY3QgdGhlIGVudHJ5J3MgbWV0aG9kb2xvZ3kgKGV4dHJhY3Rpb24gcGF0dGVybiwgY29tcHV0YXRpb24gbG9naWMsCnZlcmlmaWNhdGlvbiBzdGVwKSBhbmQgYWRhcHQgaXQgdG8gdGhlIGN1cnJlbnQgcHJvYmxlbS4KTkVWRVIgd3JpdGUgYW4gZW1wdHkgYW5ub3RhdGlvbi4KU2V0ICJ1c2VmdWwiOiBmYWxzZSBPTkxZIHdoZW4gdGhlIHRhc2sgVFlQRSBpcyBmdW5kYW1lbnRhbGx5IGRpZmZlcmVudAooY2hhcnQgcmVhZGluZyB2cy4gdGFibGUgbG9va3VwIHZzLiBuYXJyYXRpdmUgZXh0cmFjdGlvbikuCk90aGVyd2lzZSwgdXNlZnVsIE1VU1QgYmUgdHJ1ZS4KCjxHZW5lcmFsaXphdGlvbl9SdWxlPgpNYXAgdGhlIG1lbW9yeSdzIE1FVEhPRE9MT0dZLCBub3QgaXRzIGxpdGVyYWwgZW50aXR5IG5hbWVzLgpNZW1vcnk6IHNwZWNpZmljIHllYXJzL2VudGl0eSBuYW1lcyAtPiBjdXJyZW50IHByb2JsZW06IHJlbGV2YW50IGVxdWl2YWxlbnRzLgpJZiB5b3UgY2Fubm90IG1hcCBlbnRpdGllcywgdGhlIGVudHJ5IGlzIE5PVCB1c2VmdWwuCgo8T3V0cHV0PgpbeyJpIjogMCwgInVzZWZ1bCI6IHRydWUsCiAgImFubm90YXRpb24iOiAiTVVTVCBleHRyYWN0IHZhbHVlcyBmcm9tIENvbnRleHQgdGFibGUgYXMgcHl0aG9uCiAgIHZhcmlhYmxlcyBiZWZvcmUgY29tcHV0aW5nLiBETyBOT1QgY29tcHV0ZSB3aXRob3V0IHZlcmlmeWluZyB3aGljaAogICB0YWJsZSByb3cgbWF0Y2hlcyB0aGUgcXVlc3Rpb24uIn0sIC4uLl0=)<Task\>Foreachentrywrite1\-3directivesentencesusingMUST/DONOT\.Extracttheentry’smethodology\(extractionpattern,computationlogic,verificationstep\)andadaptittothecurrentproblem\.NEVERwriteanemptyannotation\.Set"useful":falseONLYwhenthetaskTYPEisfundamentallydifferent\(chartreadingvs\.tablelookupvs\.narrativeextraction\)\.Otherwise,usefulMUSTbetrue\.<Generalization\_Rule\>Mapthememory’sMETHODOLOGY,notitsliteralentitynames\.Memory:specificyears/entitynames\-\>currentproblem:relevantequivalents\.Ifyoucannotmapentities,theentryisNOTuseful\.<Output\>\[\{"i":0,"useful":true,"annotation":"MUSTextractvaluesfromContexttableaspythonvariablesbeforecomputing\.DONOTcomputewithoutverifyingwhichtablerowmatchesthequestion\."\},\.\.\.\]

Memory\_Block XML injection format[⬇](data:text/plain;base64,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)<Memory\_Block\><Field\_Guide\>EachEntryisapastexperience:\-Question:theoriginalproblem\(usetojudgescenariosimilarity\)\-Answer:thecorrectanswer\(formatreferenceonly,nevercopyvalue\)\-Experience:distilledreusablerulesorganizedas:\-Findings:patternsthatsucceeded\-\>followthesestrategies\-Cautions:guardrulesfromfailures\-\>applyifconditiontriggers</Field\_Guide\><Entry\><Question\>originalproblemtext</Question\><Answer\>correctanswer</Answer\><Experience\><Findings\>reusablestrategiesfromsuccesses</Findings\><Cautions\>guardrulesfromfailures</Cautions\></Experience\></Entry\><OptOut\>IfanyentryabovedoesNOTactuallyfitthisproblem,ignoreit\.</OptOut\></Memory\_Block\>

## Appendix HLimitations

We acknowledge two primary limitations\.First, Financial Memory construction requires auxiliary LLM calls for trajectory tagging, embedding, and experience consolidation, introducing a non\-trivial upfront cost during bank initialization\. Although retrieval remains efficient after consolidation, improving sample efficiency and enabling incremental online accumulation remain important future directions\. Second, our evaluation is limited to English financial multimodal benchmarks\. Financial documents vary substantially across languages in terminology, tabular conventions, and chart structures, which may limit direct transferability of retrieved experience and tool configurations\. Extending experience\-conditioned reasoning to multilingual financial settings and broader model families remains future work\.