acceptodds
Under review as a conference paper at ICLR 2027

Code-Interactive Visual Reasoning for Financial Candlestick Understanding

Abstract

Financial candlestick understanding requires locating visual elements, aligning them across trading dates and panels, and combining observations through numerical and financial reasoning. Answer accuracy alone does not reveal whether predictions are supported by chart evidence. We introduce Code-Interactive Financial Candlestick Reasoning and present Fin-Thinker, a Qwen3.5-9B-based model that generates and executes code to acquire evidence and perform computations, using returned observations to guide subsequent reasoning. To train these capabilities, we construct FinKLine-32K with 32,184 replay-verified code-observation trajectories collected through teacher-guided student exploration, including successful solutions with and without error recovery. Our three-stage training framework combines foundational visual-code fine-tuning, financial process supervision of successful reasoning and corrective continuations, and execution-feedback reinforcement learning. For evaluation, we introduce FinKLineBench with 1,200 questions across 138 financial task types to assess answer correctness, code execution, and evidence grounding separately. On FinKLineBench, Fin-Thinker improves strict accuracy by 13.1 and 33.2 percentage points over the strongest evaluated direct-answer baseline and its base model with sandbox access, respectively. Training-stage ablations further show that financial process supervision and execution-feedback reinforcement learning yield successive improvements in both answer accuracy and GPT-5.6-judged grounded accuracy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Financial Candlestick Understanding, Code-Interactive Visual Reasoning, Vision-Language Models, Visual Grounding, Process Supervision, Multimodal Benchmarking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gGWCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies code-interactive reasoning for financial candlestick charts. It introduces FinKLine-32K, a dataset of replay-verified reasoning trajectories, FinKLineBench, a benchmark with 1,200 questions across 138 task types, and Fin-Thinker, a Qwen3.5-9B model trained with visual-code SFT, financial process supervision, and execution-feedback RL. Fin-Thinker achieves 53.2% answer accuracy and 43.1% grounded accuracy, outperforming the evaluated baselines.

Critical strengths

1.The task is well motivated. Financial chart questions often require locating evidence across dates and panels, and code interaction is a natural way to support this process.

2.The benchmark evaluates more than final-answer accuracy. Separating answer correctness, code execution, and evidence grounding makes the model’s behavior easier to diagnose.

Critical weaknesses

1.All main results are on the authors’ own benchmark. The unseen-task split is still within the same taxonomy, so it does not show transfer to independently collected financial chart benchmarks.

2.The paper does not establish that general visual tool-use or “thinking with images” ability transfers to financial charts. Stage I already includes financial chart tasks, while the base model performs worse with code access, so the gains from tool use and financial knowledge remain mixed together.

Questions for authors

1.How does Fin-Thinker perform on relevant subsets of existing financial chart benchmarks without additional training?

2.Can general visual tool-use models solve these financial chart tasks without financial process supervision, and how much do the different training stages contribute separately?

3.How much does the reference-aware teacher affect the reasoning patterns and final performance learned from FinKLine-32K?

9 comments
AuthorsOpenReview ↗

Response (Part 1)

We sincerely thank you for your constructive feedback and recognition of our task motivation and process-aware evaluation. We appreciate your suggestions on cross-benchmark generalization, training-stage contributions, and reference-aware teacher guidance, and provide further evidence below.*

W1/Q1: Is there evaluation on independent financial-vision benchmarks without additional training?

We agree that the Seen/Unseen split within FinKLineBench alone cannot establish cross-benchmark generalization.

We select a total of 100 candlestick-related objective questions from VisFinEval (Liu et al., 2025), FinChart-Bench (Shu et al., 2026), MME-Finance (Gan et al., 2025), and https://huggingface.co/datasets/qsWDFGHN/FIN-VQA A, and evaluate Fin-Thinker without additional training. This evaluation focuses on selected candlestick-related subsets of these external datasets, with all reported results referring to the same 100-question sample.

Table G1. Zero-additional-training evaluation on 100 external financial-chart questions (accuracy, %).

ModelPattern (24)Trend (32)MACD (18)Indicator (16)Price–Volume (10)Overall (100)
Qwen3.5-9B37.543.838.937.530.039.0
DeepEyesV237.534.427.831.320.032.0
GPT-5.562.568.866.762.550.064.0
Claude Opus 4.866.771.966.762.550.066.0
Fin-Thinker (Ours)58.359.472.268.860.063.0

Fin-Thinker achieves 63.0%, outperforming Qwen3.5-9B (39.0%) and DeepEyesV2 (32.0%) by 24.0 pp and 31.0 pp, respectively, while remaining close to GPT-5.5 (64.0%) and Claude Opus 4.8 (66.0%) on this sample. Among the evaluated models, Fin-Thinker achieves the highest accuracy on MACD, Indicator, and Price–Volume, but trails GPT-5.5 and Claude Opus 4.8 on Pattern and Trend. These findings provide preliminary evidence of transfer to external financial-chart questions, while the evaluation remains limited to this 100-question sample.

AuthorsOpenReview ↗

Response (Part 2)

W2/Q2: Can general visual-tool-use models solve these financial chart tasks without financial process supervision, and how much do the different training stages contribute separately?

We thank the reviewer for raising this important concern. We conduct three complementary analyses to distinguish the contributions of general visual-tool training, foundational financial chart skills, and financial process supervision: base-model evaluation with code access, Stage I data ablation, and supervision-format comparisons across training stages.

1. Base-model Performance with Code Access.

Qwen3.5-9B achieves 23.6% accuracy through direct answering, while its accuracy drops to 20.0% when code execution is enabled. This performance degradation suggests that image-grounded code reasoning in financial chart analysis is challenging for general-purpose models without task-specific training. It requires coordinating visual evidence extraction, date and indicator identification, code execution, and subsequent reasoning. Simply providing code execution access does not equip models with these capabilities, motivating our staged training approach.

2. Separating General Visual-Tool Training from Basic Financial Skills.

To disentangle the contributions of general visual-tool training and task-specific financial chart skills, we conduct a Stage I data ablation. The 6,570 Stage I financial examples are not general financial knowledge . Instead, they are some atomic-tool trajectories for reading and interpreting candlestick charts, including extracting price and volume information, locating exact trading dates, identifying chart patterns and technical indicators, detecting MA5–MA10 crossovers, and counting price-rise and volume-expansion days. These examples teach foundational operations for acquiring and interpreting financial chart evidence, rather than complete multi-step financial reasoning trajectories. We compare three Stage I configurations: General TWI Only (13,910 examples), Basic Financial Skills Only (6,570 examples), and their combination (20,480 examples). all configurations subsequently use the same FinKLine-32K data in Stage II and the full execution-feedback reward in Stage III.

Table G2. Stage I data ablation on FinKLineBench (accuracy, %).

Stage I dataStage IAfter Stage IIAfter Stage III
General TWI Only22.442.545.2
Basic Financial Skills Only28.445.750.8
General TWI + Basic Financial Skills (Ours)36.449.153.2

Adding basic financial skills to general TWI training raises Stage I accuracy from 22.4% to 36.4%. After all three stages, the financial-skills-only configuration reaches 50.8%, compared with 45.2% for the general-only configuration, while combining both sources achieves 53.2%. These results demonstrate complementary contributions from general visual-tool training and foundational financial chart skills. The substantial improvements after Stage II further demonstrate the additional value of full financial process supervision from FinKLine-32K.

3. Separating Supervision Formats Across Training Stages.

We further compare Answer-only, Textual CoT, Code (6K), and Code (Full) across training stages. All variants are based on Qwen3.5-9B, whose base accuracy on FinKLineBench is 23.6%.

In Stage I, all variants use the same 6,570 foundational financial chart tasks but differ in supervision: Answer-only uses final-answer supervision, Textual CoT uses GPT-distilled reasoning, and Code (6K) uses corresponding atomic-tool trajectories to supervise chart-evidence operations, some without final answers. Code (Full) further adds 13,910 public TWI trajectories. In Stage II, all variants are trained on the same 32,184 FinKLine-32K instances under their respective supervision formats.

Table G3. Accuracy across training stages on FinKLineBench (additional evaluation, %).

Training stageAnswer-onlyTextual CoTCode (6K)Code (Full)
Stage II only (without Stage I)29.234.6——
Stage I25.026.328.436.4
Stage I+II30.434.645.749.1
Stage I + II + III (shared answer reward)31.136.748.751.7
Stage I + II + III (full execution-feedback reward)——50.853.2

Q3: What does the reference-aware teacher contribute to trajectory construction and final student performance?

We compare reference-free and reference-aware teacher guidance on the same 500 tasks, including 200 hard tasks, separately from the 690-task experiment in paper Table 5. The intervention is whether the teacher can access reference answers during offline correction.

Table G3. Reference-aware teacher ablation: trajectory construction.

MetricReference-freeReference-aware
Acceptance ↑24.0% (120/500)73.0% (365/500)
Hard-task acceptance ↑7.0% (14/200)59.0% (118/200)
Correction rounds ↓3.62.1

The reference-aware teacher helps turn unsuccessful explorations into verified training trajectories, providing supervision on evidence acquisition and error recovery. On the same 500 tasks, reference access increases acceptance from 24.0% to 73.0%, and hard-task acceptance from 7.0% to 59.0%. The student generates the operations and corrections; teacher feedback and reference-answer inputs are excluded, and retained executions undergo independent replay.

Complementary evidence from the recovery-supervision ablation shows an accuracy improvement from 44.3% to 49.1% (paper Table 4C), supporting the downstream value of corrective trajectories.

Reviewer_9HHLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a model named Fin-Thinker, which uses executable code to interact with financial K-line charts and perform multi-step reasoning. The authors introduce FinKLineBench containing 1200 questions and FinKLine-32K containing 32184 verified trajectories. The model is trained through visual-code SFT, financial process supervision, and execution-feedback RL. The results show that, compared with multiple baselines, the answer accuracy, code execution, and evidence grounding all have improvements.

Critical strengths

1.I think the dataset and benchmark are valuable contributions. The paper covers a wide range of candlestick tasks and provides detailed code-execution trajectories, including error recovery. Evaluating evidence grounding in addition to final-answer accuracy is also useful.

2.The experimental results are encouraging. Fin-Thinker achieves clear improvements over the evaluated baselines, and the training-stage and recovery-supervision ablations provide useful evidence that the proposed training pipeline is effective.

Critical weaknesses
  1. I am not very convinced by the methodological novelty. Code-based visual reasoning, tool interaction, and program-guided supervision have already been explored by previous work, such as ViperGPT, Visual Program Distillation, and ChartAgent. The main contribution here seems to be to adapt these ideas to financial candlestick charts, plus a new dataset and a three-stage training pipeline. I think if the authors can better isolate what is new in the proposed method, the paper will be stronger. Especially, this point is not very clear to me: how much improvement is from code interaction itself, not from domain-specific training data and additional optimization.

  2. Some experimental comparisons do not fully isolate the claimed contributions. The comparison with Answer-only and Textual Reasoning SFT is not very controlled. As clarified in Appendix F.1, these baselines do not receive Stage I training or Stage III RL, which is different from the full model. Therefore, the large performance gap cannot be attributed only to code-interactive reasoning. I also have concerns about how much the benchmark supports the broader reasoning claims. Although the overall accuracy reaches 53.2%, the performance on compositional reasoning is still only 14.3%. Moreover, the paper does not clearly report the separate quantitative results for synthetic and real-world charts. I would like to see stronger evidence that the learned capabilities can generalize beyond the benchmark's rendering styles and closely related task types.

Questions for authors
  1. What is the main methodological difference between your method and existing code-based visual reasoning methods? Can you compare with a stronger code-interactive baseline that is trained on the same financial data, not mainly compare with pretrained interactive models? 

  2. Can you give a more controlled comparison between code-interactive reasoning and textual reasoning? Especially, can the textual baseline receive similar training stages and optimization budgets? This will help clarify whether the improvement is from code interaction or from additional training. 

  3. Can you give more analysis on compositional reasoning? Now the accuracy is only 14.3%, although the paper emphasizes multi-step financial reasoning. What are the main failure points? And how much improvement is from better visual extraction, and how much is from downstream reasoning?

6 comments
AuthorsOpenReview ↗

Response (Part 1)

We sincerely thank you for your thoughtful review and recognition of the value of our dataset and benchmark. We appreciate your concerns regarding methodological novelty, comparison fairness, and compositional reasoning, and provide additional experiments and clarifications below.

W1/Q1: What distinguishes the method from prior code-based visual reasoning, and why not use a same-data code baseline?

We acknowledge prior code-based visual reasoning methods, including ViperGPT, Visual Program Distillation (VPD), and ChartAgent. Our contribution is a replay-verified process-supervision pipeline tailored to financial visual reasoning, addressing temporal localization, cross-panel alignment, and error propagation through financial atomic tools, student exploration, reference-aware teacher correction, recovery supervision, and independent replay.

Table H1. Methodological comparison with prior work.

MethodTrainingInferenceCodeMulti-turnFeedbackRecoverySFTClean-stateReplayProcessDiagnostics
ViperGPTNoneYesNoNoNoProgram traces
VPDSFTNoNoNoNoHuman factuality
ChartAgentNoneTool APIsYesNoNoSelf-verification
DeepEyesRLYesYesNoNoTool analysis
ThymeSFT + RLYesYesNoNoCode analysis
Fin-ThinkerSFT + GRPOYesYesYesYesER / GAcc / DR

Recovery SFT supervises corrective trajectories, while clean-state replay verifies their independent re-execution. Our construction ablations further support these design choices: the complete pipeline increases valid-trajectory yield from 44.9% to 74.8% over direct teacher generation; financial atomic tools raise it from 55.5% to 74.8% over generic code tools (paper Table 5); and recovery supervision improves downstream accuracy from 44.3% to 49.1% (paper Table 4C). To examine whether our financial process supervision transfers to another model, we further train DeepEyesV2 using Financial 6K and FinKLine-32K, followed by execution-feedback RL.

Table H2. Additional training comparisons on FinKLineBench (accuracy, %).

MethodFinancial6KSFTFinKLine-32KSFTRLRewardBeforeRL45Steps90Steps
DeepEyesV2 (Original)NoNo—16.2——
DeepEyesV2 + GRPO (Original Reward)NoNoDeepEyesV216.221.621.2
DeepEyesV2 + GRPO (Our Reward)NoNoOurs16.222.524.6
DeepEyesV2 + FinKLine-32K SFTNoYes—22.4——
DeepEyesV2 + Financial 6K + FinKLine-32K SFTYesYes—27.6——
DeepEyesV2 + Financial 6K + FinKLine-32K SFT + RLYesYesOurs27.630.834.5
Fin-Thinker (Ours)YesYesOurs49.153.255.8

Financial 6K contains 6,570 atomic-tool trajectories. The two direct-GRPO variants share the same initialization and training setup, differing only in reward. We extend training to 90 steps to examine further gains; the separate training-stage comparison in Table H6 uses 45 steps.

Direct GRPO with the original reward shows no further improvement after 45 steps (21.6% → 21.2%), while our reward reaches 24.6% at 90 steps. Adding FinKLine-32K SFT raises DeepEyesV2 to 22.4%, and incorporating Financial 6K preparation further improves it to 27.6%. Subsequent RL reaches 34.5%, outperforming direct GRPO with the same reward and training budget by 9.9 pp.

These results highlight the importance of foundational financial operations for learning complex financial reasoning trajectories. Financial 6K prepares the model for precise numerical extraction, trading-day localization, and cross-panel alignment, while FinKLine-32K supervises how these operations are integrated into multi-step reasoning. The improvements on DeepEyesV2 demonstrate that our financial process supervision also benefits another model. Fin-Thinker additionally incorporates public TWI data in Stage I and reaches 55.8% at 90 steps, further supporting the value of the staged approach.

AuthorsOpenReview ↗

Response (Part 2)

Additional Analysis on Generalization Across Chart Types and Sources

We further evaluate whether Fin-Thinker generalizes beyond the training task types and chart rendering styles.

1. Seen vs. Unseen Chart Types and Reasoning Tasks FinKLineBench contains 28 visual chart types, of which 16 are covered during training. Fin-Thinker achieves 55.1% accuracy on seen chart types and 47.3% on unseen chart types.

Table H3. Chart-type exposure analysis.

CharttypesNumberoftypesQuestionsAccuracy(%)
Seen in training1690055.1
Unseen in training1230047.3
Overall281,20053.2

For reasoning task types, Fin-Thinker achieves 55.32% on seen tasks and 50.47% on unseen tasks, with a gap of only 4.85 pp.

2. Synthetic vs. Real-world Charts We separately evaluate performance on programmatically rendered charts and real-world screenshots.

Table H4. Chart-source performance by reasoning category.

ReasoningcategorySynthetic(n)Syntheticacc.(%)Real(n)Realacc.(%)
VP7598.75796.5
TCR31054.823447.9
FP24551.020944.0
CR3016.74012.5
Overall66056.754048.9

Fin-Thinker achieves 56.7% on synthetic charts and 48.9% on real-world screenshots, with a 7.8 pp gap, demonstrating generalization across different chart sources.

3. Generalization to Independent Financial-chart Benchmarks

Finally, we evaluate Fin-Thinker on 100 externally sourced financial-chart questions without additional training.

Table H5. Zero-additional-training evaluation on external financial-chart questions (accuracy, %).

ModelPattern(24)Trend(32)MACD(18)Indicator(16)Price–Volume(10)Overall
Qwen3.5-9B37.543.838.937.530.039.0
DeepEyesV237.534.427.831.320.032.0
GPT-5.562.568.866.762.550.064.0
Claude Opus 4.866.771.966.762.550.066.0
Fin-Thinker (Ours)58.359.472.268.860.063.0

Fin-Thinker achieves 63.0%, substantially outperforming Qwen3.5-9B (39.0%) and DeepEyesV2 (32.0%), while approaching GPT-5.5 (64.0%) and Claude Opus 4.8 (66.0%). These results provide further evidence of its generalization to independently sourced financial-chart questions.

AuthorsOpenReview ↗

Response (Part 3)

W2a/Q2: Are the textual baselines fairly matched in training stages and optimization budget?

We extend the textual baselines to all three training stages and add a Code (6K) variant based on the same financial seed-task pool. We compare supervision formats and report the corresponding training budgets. All variants build on Qwen3.5-9B, whose base accuracy on FinKLineBench is 23.6%.

Training configurations. In Stage I, Answer-only and Textual CoT use 6,570 financial QA examples, with final-answer supervision and GPT-distilled reasoning, respectively. Code (6K) uses 6,570 financial atomic-tool trajectories constructed from the same seed-task pool; some contain intermediate operations without a final answer. Code (Full) additionally uses 13,910 public TWI trajectories. In Stage II, all variants use the same 32,184 financial QA instances with their corresponding supervision formats. In Stage III, the comparison uses 45 GRPO steps, four rollouts per prompt, and the same answer-correctness reward. We separately report the code variants trained with the full execution-feedback reward.

Table H6. Accuracy across training stages on FinKLineBench (additional evaluation, %).

Training stageAnswer-onlyTextual CoTCode (6K)Code (Full)
Stage II only (without Stage I)29.234.6——
Stage I25.026.328.436.4
Stage I+II30.434.645.749.1
Stage I + II + III (shared answer reward)31.136.748.751.7
Stage I + II + III (full execution-feedback reward)——50.853.2

Code (6K) achieves 28.4% after Stage I, where supervision primarily targets atomic financial tool operations and includes intermediate steps without final answers. Code (Full) additionally uses 13,910 public TWI trajectories that provide complete end-to-end reasoning with final-answer supervision. These trajectories provide complete sequences of tool operations, returned observations, reasoning, and final answers, complementing the atomic financial-operation supervision in Code (6K). With this combined supervision, Code (Full) reaches 36.4% after Stage I.

Stage II further trains the model to integrate financial operations into complete evidence acquisition and reasoning trajectories. Code (6K) then reaches 45.7%, compared with 34.6% for Textual CoT. With the shared answer reward, Code (6K) further improves to 48.7%, exceeding Textual CoT’s 36.7% by 12.0 percentage points (pp) and the base model’s 23.6% by 25.1 pp. With the full execution-feedback reward, Code (6K) and Code (Full) reach 50.8% and 53.2%, respectively.

Table H7. Training-budget estimates and inference costs.

Budget / costAnswer-onlyTextual CoTCode (6K)Code (Full)
Stage I samples × epochs6,570 × 16,570 × 16,570 × 120,480 × 1
Stage II samples × epochs32,184 × 132,184 × 132,184 × 132,184 × 1
SFT supervised tokens (M)0.413.7553.659.4
RL steps45454545
Rollouts / prompt4444
Avg. generated tokens / rollout1955683,0004,933
RL rollout tokens (M)0.280.824.327.10
Avg. inference tokens / QA1328293,8223,980
Avg. code turns / QA——4.54.7

We report training-token counts and inference costs based on our actual experimental runs. For code variants, we measure inference costs under the full execution-feedback configuration.

What the additional supervision enables. Code (6K) receives 53.6M supervised SFT tokens, compared with 3.75M for Textual CoT. The code trajectories supervise chart inspection; the extraction of prices, volumes, and technical indicators; alignment across trading dates and panels; and reasoning over the resulting evidence. These trajectories provide executable examples of integrating financial knowledge, intermediate evidence, and numerical computation to solve multi-step financial questions. These results support the value of executable process supervision under the reported training budgets.

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

This paper introduces Fin-Thinker, a model trained from Qwen3.5-9B that answers questions about candlestick charts by writing and running code to crop, measure, and compute. It is trained in three stages: visual-code SFT, SFT on replay-verified trajectories collected with a GPT-5.6 teacher, and GRPO with execution-based rewards. The paper also introduces FinKLineBench, which scores answer accuracy, code execution rate, and whether each correct answer is grounded in chart evidence. Fin-Thinker reaches 53.2% accuracy, compared with 40.1% for the best direct-answer model (Claude Opus 4.8).

Critical strengths
  • Complete and verified data pipeline: Every training trajectory is replayed in a clean sandbox before being kept, and the ablation shows that including error-recovery trajectories helps.
  • Process-level evaluation: The benchmark checks not only the final answer but also whether it is supported by evidence from the correct date and panel. This process-level evaluation is important for chart reasoning, where a correct answer can come from the wrong evidence.
Critical weaknesses
  • Training and evaluation are closely coupled: The training pool is built with the same annotation protocol and chart sources as FinKLineBench. The results are not broken down by synthetic versus real charts. It is therefore hard to tell how much of the gain reflects fit to this data distribution rather than general chart-reasoning ability.
  • Limited methodological novelty: The training recipe closely follows existing thinking-with-images work such as DeepEyes, Mini-o3, and Thyme. The main difference is the application of an established pipeline to a single narrow domain. The main new contribution is the benchmark rather than the method.
Questions for authors
  • What is the accuracy on real screenshots versus programmatically rendered charts?
  • Has the hallucination judge been validated against human labels?
3 comments
AuthorsOpenReview ↗

Response (Part 1)

We sincerely thank you for your constructive feedback and recognition of our trajectory construction and process-aware evaluation. We appreciate your suggestions on generalization, methodological novelty, and evaluation reliability, and provide additional experiments and clarifications below.

W1/Q1: How well does the method generalize across task types and real versus synthetic charts?

We evaluate generalization from three perspectives: reasoning task types, visual chart types, and chart sources (synthetic versus real-world screenshots). We further assess transfer to independently collected financial-chart questions.

Data separation and diversity. FinKLine-32K and FinKLineBench share the same annotation protocol and chart-source categories. However, we prevent training–evaluation overlap through stock-ticker and image-hash deduplication, together with checks for overlapping trading periods and re-rendered charts. FinKLine-32K contains 32,184 verified trajectories derived from 18,742 images, including 12,370 real screenshots (66.0%) and 6,372 synthetic charts (34.0%). Real charts are captured from platforms such as TradingView, whereas synthetic charts are programmatically rendered from real OHLCV data.

Generalization across reasoning tasks and chart types. FinKLineBench contains 1,200 questions across 864 distinct images, covering 28 visual chart types and 138 reasoning task types (paper Table 10). The training data cover 16 visual chart types. For reasoning task generalization, Fin-Thinker achieves 55.32% accuracy on seen task types and 50.47% on unseen task types, a gap of only 4.85 percentage points (pp; paper Table 2). We further evaluate generalization across seen and unseen visual chart types.

Table P1. Generalization across seen and unseen chart types.

Chart typesNumber of typesQuestionsAccuracy (%)
Seen in training1690055.1
Unseen in training1230047.3
Overall281,20053.2

Fin-Thinker retains 47.3% accuracy on the 12 chart types absent from training, demonstrating its ability to generalize beyond the visual chart types encountered during training.

Generalization across real and synthetic charts. We further break down performance by chart source and reasoning category to examine whether the model is sensitive to chart rendering styles.

Table P2. Performance on synthetic and real-world charts.

Reasoning categorySynthetic (n)Synthetic acc. (%)Real (n)Real acc. (%)
VP7598.75796.5
TCR31054.823447.9
FP24551.020944.0
CR3016.74012.5
Overall66056.754048.9

Fin-Thinker achieves 56.7% accuracy on synthetic charts and 48.9% on real screenshots, with a 7.8 pp gap. Although performance is lower on real screenshots, the model maintains substantial accuracy across both sources, providing evidence of generalization beyond programmatically rendered charts.

Generalization to independent financial-chart benchmarks. Finally, to assess transfer beyond FinKLineBench, we evaluate Fin-Thinker on 100 externally sourced candlestick-related questions without additional training. The questions are selected from VisFinEval (Liu et al., 2025), FinChart-Bench (Shu et al., 2026), MME-Finance (Gan et al., 2025), and FIN-VQA. All models are evaluated on the same 100-question sample.

Table P3. Zero-additional-training evaluation on external financial-chart questions (accuracy, %).

ModelPattern (24)Trend (32)MACD (18)Indicator (16)Price–Volume (10)Overall
Qwen3.5-9B37.543.838.937.530.039.0
DeepEyesV237.534.427.831.320.032.0
GPT-5.562.568.866.762.550.064.0
Claude Opus 4.866.771.966.762.550.066.0
Fin-Thinker (Ours)58.359.472.268.860.063.0

Fin-Thinker achieves 63.0% accuracy, outperforming Qwen3.5-9B and DeepEyesV2 by 24.0 pp and 31.0 pp, respectively. It also remains competitive with GPT-5.5 (64.0%) and Claude Opus 4.8 (66.0%), achieving the highest accuracy among the evaluated models on MACD, technical indicators, and price–volume reasoning. These results provide additional evidence of transfer to independently sourced financial-chart questions.

AuthorsOpenReview ↗

Response (Part 2)

W2: Is this mainly an application of existing Thinking-with-Images methods to finance?

Existing Thinking-with-Images (TWI) methods, including DeepEyes, Mini-o3, and Thyme, have advanced tool-assisted visual reasoning. However, the TWI models evaluated in paper Table 1 achieve only 1.2–18.3% accuracy on FinKLineBench. The process-level evaluation in paper Table 3 reports execution rates of 55.2–63.5% and grounded accuracy of 5.1–8.4%. These findings highlight the gap between executable tool use and evidence-grounded financial reasoning, where precise numerical extraction, temporal localization, and cross-panel alignment remain challenging.

Our work addresses an underexplored gap in financial visual reasoning by identifying the limitations of existing Thinking-with-Images (TWI) methods in precise numerical extraction, evidence grounding, and multi-step reasoning. To address these challenges, we propose an efficient, replay-verified trajectory synthesis framework that integrates financial atomic tools, teacher-guided student exploration, and error-recovery supervision. Beyond candlestick charts, the framework is designed to be extensible to other financial visual documents, such as PDF financial reports, where numerical accuracy and verifiable reasoning are essential. Unlike prior methods that primarily focus on visual tool use, program generation, or general-purpose visual reasoning, Fin-Thinker emphasizes verifiable financial process supervision.

Table P4. Methodological comparison with prior work.

MethodTrainingInference CodeMulti-turn FeedbackRecovery SFTClean-state ReplayProcess Diagnostics
ViperGPTNoneYesNoNoNoProgram traces
VPDSFTNoNoNoNoHuman factuality
ChartAgentNoneTool APIsYesNoNoSelf-verification
DeepEyesRLYesYesNoNoTool analysis
ThymeSFT + RLYesYesNoNoCode analysis
Fin-ThinkerSFT + GRPOYesYesYesYesER / GAcc / DR

Our ablations support these design choices: In our construction ablations, the full pipeline achieves a valid-trajectory yield of 74.8%, compared with 44.9% for direct teacher generation and 55.5% for the configuration using generic code tools (paper Table 5). Recovery supervision also increases downstream accuracy from 44.3% to 49.1% (paper Table 4C). Across training stages, ER increases from 82.4% to 87.2%, while GAcc rises from 22.1% to 43.1% (Appendix F.1, paper Table 19), showing improvements in evidence grounding alongside more reliable execution.

Q2: Has the hallucination / grounding judgment been validated against human annotations?

The grounding evaluator was previously validated through a human audit of 300 samples (Sec. 6.3; App. A.5), achieving 97.0% agreement and Cohen’s κ = 0.938.

To further validate the hallucination evaluator, we conducted a stratified, blinded human study of 100 trajectories, sampling 20 from each of five models: Fin-Thinker, Claude Opus 4.8, DeepEyesV2, Qwen3.5-9B Direct, and Qwen3.5-9B + Code. Five annotators independently assessed the trajectories without access to model identities or evaluator predictions. Majority voting identified 34 hallucination cases (34.0%).

Table P5. Human validation of the hallucination evaluator.

MetricResult
Inter-annotator agreement (Fleiss’ κ)0.80
Human–evaluator agreement93.0%
Cohen’s κ0.85
Precision / Recall / F188.6% / 91.2% / 89.9%

These results demonstrate strong agreement between human judgments and the hallucination evaluator.

AuthorsOpenReview ↗

Kind Reminder and Follow-up

Dear Reviewer PfjH,

Thank you again for your constructive review. We would like to kindly follow up on our rebuttal, in which we provided additional evaluations on real versus synthetic charts, unseen chart types, and external financial benchmarks, clarified our methodological contributions beyond existing Thinking-with-Images approaches, and supplemented human validation of the hallucination evaluator.

As the discussion period is approaching its end, we would greatly appreciate knowing whether these additional experiments and clarifications have addressed your main concerns regarding generalization, methodological novelty, and evaluation reliability. If any questions remain, we would be happy to provide further clarification.

Thank you again for your time and consideration.

Comments on the paper

3 comments
Area_Chair_xjH8OpenReview ↗

Discussion between Authors and Reviewers

Dear Reviewers,

The authors have submitted their rebuttal. Please read it and discuss with the authors via posted messages if necessary.

Note that the author-reviewer discussion phase ends on November 18 at 11:59 pm AoE. It is appreciated if you could share your thoughts as soon as possible, so the authors have enough time to clarify any remaining concerns.

Thank you!

Best, Your AC

AuthorsOpenReview ↗

Summary of Contributions and Significance

We sincerely thank all reviewers for their constructive feedback. In response to the concern of "limited methodological novelty" raised by multiple reviewers, we would like to further clarify the positioning of our contributions. Our goal is neither to introduce a new LLM architecture or a standalone supervision algorithm, nor to simply adapt existing Thinking-with-Images (TWI) methods to the financial domain. Instead, using financial candlestick charts as a representative setting, we investigate how to systematically develop and improve agentic reasoning capabilities for complex financial visual tasks, establishing a comprehensive research paradigm that spans problem formulation, verifiable data construction, model training, and process-aware evaluation.

This research is motivated by the distinctive challenges of information-dense financial visuals. Models must not only acquire precise visual and numerical evidence through code interaction, but also establish temporal, spatial, and logical relationships across observations and integrate them through multi-step reasoning to produce reliable answers. Our systematic evaluation reveals substantial limitations of existing general-purpose TWI methods in these capabilities. Moreover, constructing long-horizon agentic training trajectories involving complex evidence dependencies, tool feedback, and error recovery is difficult to achieve reliably through direct teacher distillation.

To address these challenges, our contributions comprise four interconnected components:

  1. Problem Formulation: We introduce Code-Interactive Financial Candlestick Reasoning, which integrates code-driven evidence acquisition, cross-evidence reasoning, and financial computation. Through systematic evaluation, we identify the limitations of existing VLMs and TWI methods in complex financial visual reasoning.

  2. Data Construction: We develop a framework for synthesizing verifiable, long-horizon agentic trajectories by combining iterative refinement of financial atomic tools, teacher-guided student exploration, error-recovery supervision, and independent execution replay. This framework produces FinKLine-32K, comprising 32,184 verified trajectories that provide reliable process-level supervision for complex financial visual reasoning. Upon acceptance, we will publicly release all training data, evaluation benchmarks, and data synthesis code to the research community.

  3. Model Training: We train Fin-Thinker through three stages: foundational visual-code supervision, financial process supervision, and execution-feedback reinforcement learning. This progressive training enables the model to generate and reuse intermediate visual evidence and computational results, establish relationships across observations, and perform complex financial reasoning.

  4. Evaluation: We introduce FinKLineBench, comprising 1,200 questions across 138 task types, and systematically evaluate open-source and proprietary models under both direct-answer and interactive reasoning settings. Beyond answer correctness, our evaluation examines code execution reliability, evidence grounding, and the reproducibility of execution dependencies, providing a more diagnostic assessment of financial visual reasoning capabilities.

In summary, our central contribution is not an isolated model architecture, training algorithm, or evaluation dataset, but a unified methodological framework that connects problem formulation, verifiable agentic data synthesis, progressive capability training, and process-aware evaluation for complex financial visual reasoning. Rather than merely assessing what existing VLMs can accomplish in financial tasks, we focus on how to construct reliable training data, enable models to learn verifiable evidence-based reasoning processes, and systematically improve their ability to solve complex financial problems.

AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair

Dear Area Chair,

Thank you for your time and effort in handling our submission. As the discussion period has concluded without further responses from the reviewers, we summarize how our rebuttal and additional experiments address the main concerns raised by Reviewers PfjH (3), 9HHL (2), and gGWC (2).

1. Generalization (PfjH W1/Q1, 9HHL W2b, gGWC W1/Q1)

  • External Benchmarks: Evaluated without additional training on 100 candlestick questions from four independent datasets (VisFinEval, FinChart-Bench, MME-Finance, FIN-VQA). Fin-Thinker reaches 63.0% accuracy, outperforming Qwen3.5-9B (39.0%) and DeepEyesV2 (32.0%), approaching GPT-5.5 (64.0%) and Claude Opus 4.8 (66.0%), and ranking first on MACD (72.2%), Indicators (68.8%), and Price-Volume (60.0%).
  • Real vs. Synthetic Charts: On FinKLineBench, Fin-Thinker achieves 56.7% on synthetic charts (660) and 48.9% on real screenshots (540), a 7.8 pp gap.
  • Unseen Types: Fin-Thinker retains 47.3% accuracy on 12 unseen visual chart types (vs. 55.1% on 16 seen types) and 50.47% on unseen reasoning tasks (vs. 55.32% on seen tasks).

2. Matched Comparisons & Component Analysis (9HHL W1/W2a/Q1/Q2, gGWC W2/Q2)

  • 3-Stage Textual Baselines: We extended Answer-only and Textual CoT to all three stages using the exact same financial data and 45 GRPO steps (shared answer reward). Code (6K) reaches 48.7%, outperforming Textual CoT (36.7%) and Answer-only (31.1%). Adding our full execution-feedback reward further lifts Code (6K) to 50.8% and Code (Full) to 53.2%. Full training-token and inference costs are also reported.
  • Stage I Data: After all three stages, General TWI Only reaches 45.2%, Basic Financial Skills Only reaches 50.8%, and combining both achieves 53.2%, showing clear complementary gains.
  • Transfer to DeepEyesV2: Applying our Financial 6K + FinKLine-32K SFT and execution-feedback RL improves DeepEyesV2 from 16.2% to 34.5% (at 90 steps), outperforming direct GRPO with its original reward (21.2%) or our reward (24.6%).

3. Methodological Novelty & Data Synthesis (PfjH W2, 9HHL W1, gGWC Q3)

  • Beyond Existing TWI: General-purpose TWI models (DeepEyes, Mini-o3, Thyme) achieve only 1.2% to 18.3% accuracy and 5.1% to 8.4% grounded accuracy on FinKLineBench. Rather than a simple domain adaptation, our work introduces a unified, replay-verified process-supervision framework combining financial atomic tools, student exploration, reference-aware error-recovery SFT, and clean-state replay.
  • Reference-Aware Teacher: On 500 tasks (200 hard), reference-aware correction increases trajectory acceptance from 24.0% to 73.0% (hard tasks: 7.0% to 59.0%) and reduces correction rounds from 3.6 to 2.1. Recovery SFT boosts downstream accuracy from 44.3% to 49.1%.

4. Compositional Reasoning Analysis (9HHL W2b/Q3)

  • Error Breakdown (70 cases): Primary bottlenecks are trading-day/cross-panel alignment (34.3%) and visual numerical extraction (21.4%), followed by event interpretation (17.1%), aggregation (8.6%), and code execution (4.3%).
  • Oracle Evidence: Providing verified dates, OHLCV values, and indicators lifts Fin-Thinker from 14.3% (10/70) to 41.4% (29/70), surpassing Textual CoT at 27.1% (19/70).

5. Human Validation of Evaluators (PfjH Q2)

  • Alongside our prior grounding-evaluator audit (97.0% agreement, Cohen's κ = 0.938), a blinded 5-annotator study on 100 trajectories for the hallucination evaluator achieved Fleiss' κ = 0.80, 93.0% agreement, Cohen's κ = 0.85, and 89.9% F1.

All reviewers recognized the value of our process-level benchmark, and PfjH and 9HHL further highlighted our verified trajectory pipeline. We believe these additional experiments, external evaluations, and human validations directly address all remaining soundness and generalization concerns. Upon acceptance, we will publicly release FinKLine-32K, FinKLineBench, and the full data synthesis and evaluation code.

Thank you again for your careful consideration.

Best regards,
Authors of Submission 8008

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.