BlueFin: A Multi-Disciplinary Overview

• BlueFin is a multi-context term encompassing an LLM benchmark for financial spreadsheets, underwater robotics design, robust world-model learning, and biomimetic marine studies.
• Underwater robotics research leverages BlueFin-inspired systems for fin-actuated propulsion and ambient current estimation using advanced simulation, control, and sensor integration techniques.
• The financial spreadsheet benchmark evaluates LLM agents on synthesis, manipulation, and interrogation tasks with granular criteria, while biomimetic studies draw on bluefin tuna for design and economic insights.

Searching arXiv for papers related to "BlueFin" and closely associated usages. BlueFin is a label that appears in several distinct technical contexts rather than as a single canonical object. In the cited literature, it most explicitly names a benchmark for evaluating LLM agents on professional financial spreadsheets, while cognate forms also designate BlueFin-like underwater robots, the Bluefin-21 unmanned underwater vehicle, the Bluefin-Tuna software organization, and the biological referent bluefin tuna that motivates both ecological theory and biomimetic robot design (Kundurthy et al., 29 May 2026, Hamamatsu et al., 5 Feb 2025, Wolek et al., 2024, Zollicoffer et al., 30 Nov 2025, Burgess et al., 2014).

1. Terminological scope

In current research usage, the term spans AI evaluation, underwater robotics, autonomy, and fisheries-related science. The exact referent is domain-specific (Kundurthy et al., 29 May 2026, Wolek et al., 2024, Zollicoffer et al., 30 Nov 2025).

Usage	Research role
BlueFin	Benchmark for LLM agents on financial spreadsheets
BlueFin-like robot	Reference frame for fin-actuated underwater robot design and control
Bluefin-21	Unmanned underwater vehicle used in current-estimation experiments
Bluefin-Tuna	Organization releasing the WISER robustness codebase
bluefin tuna	Biological and economic subject, and biomimetic template

This multiplicity matters because superficially similar names correspond to technically unrelated artifacts. In one branch of the literature, BlueFin is an evaluation harness with rubrics, judge models, and spreadsheet tools; in another, it denotes or evokes fin-propelled underwater systems whose central problems are hydrodynamics, force control, and navigation in flow; in a third, it appears through the species name bluefin tuna, which anchors both conservation theory and robot morphology (Kundurthy et al., 29 May 2026, Hamamatsu et al., 5 Feb 2025, Burgess et al., 2014).

2. BlueFin-like underwater robots and fin-actuated propulsion

Several underwater-robotics papers discuss methods explicitly framed as relevant to a BlueFin-like robot. One study develops a practical Real2Sim2Real pipeline for a soft-fin actuator representative of a multi-fin robot with four individually actuated soft flippers. The actuator is a soft silicone fin made from Zhermack Elite Double 22, cast around a short aluminum bracket and attached to a Maxon EC-max 30 brushless motor with Hall encoder, driven by an EPOS2 Module 36/2 digital positioning controller. The control objective is not scalar thrust alone but two-dimensional force-vector generation in the fin frame, $(F_x,F_y)$, from motor commands $\mathcal{A}_t=(\theta_t,\omega_t)$. A still-water dataset of 100 logs, each with 20,000 sensor readings at 100 Hz, yields 2,000,000 training examples. The learned surrogate is decomposed into PosNet, a 1D CNN with window $100$ and $113{,}601$ parameters, and ForceNet, an LSTM-based model with window $100$ and $129{,}794$ parameters. Average held-out accuracy is reported as RMSE $0.03$, MAE $0.002$, DTW $0.002$ for PosNet and RMSE $0.2$, MAE $\mathcal{A}_t=(\theta_t,\omega_t)$0, DTW $\mathcal{A}_t=(\theta_t,\omega_t)$1 for ForceNet. Training then proceeds with LSTM-PPO; a single generalized controller uses 30,000 training steps, whereas each grid-switching model uses 10,000 steps. In real deployment over six force references, grid-switching RL reduces overall mean $\mathcal{A}_t=(\theta_t,\omega_t)$2 error from $\mathcal{A}_t=(\theta_t,\omega_t)$3 to $\mathcal{A}_t=(\theta_t,\omega_t)$4 and overall mean $\mathcal{A}_t=(\theta_t,\omega_t)$5 error from $\mathcal{A}_t=(\theta_t,\omega_t)$6 to $\mathcal{A}_t=(\theta_t,\omega_t)$7; one model occupies about 7.2 MB (Hamamatsu et al., 5 Feb 2025).

A broader hydrodynamic treatment models fin-based propulsion with a NACA 0020 hydrofoil in a uniform freestream using WaterLily, a two-dimensional incompressible solver based on the Boundary Data Immersion Method. The canonical motion is combined heave and leading-edge pitch, with baseline operation at $\mathcal{A}_t=(\theta_t,\omega_t)$8, $\mathcal{A}_t=(\theta_t,\omega_t)$9, and $100$0. The reported single-fin baseline is $100$1 and $100$2. A reduced-order flexibility model with a leading-edge torsional spring and time-varying stiffness improves these values to $100$3 and $100$4, corresponding to an 8% increase in thrust and a 6.4% increase in efficiency. In tandem-fin interaction studies, the best sampled two-fin case is $100$5, $100$6, with $100$7, whereas a poor timing choice at $100$8, $100$9 reduces that ratio to $113{,}601$0 (Grobe, 24 Apr 2026).

Biomimetic median-fin studies sharpen the same design picture. A free-swimming tuna-inspired robotic fish with a magnetically attached morphing dorsal fin finds that the erected dorsal fin substantially improves yaw stability while showing little influence on mean speed and cost of transport over the tested range. Peak-to-peak yaw reduction ranges from 16.51% to 24.20%, with the largest improvement at $113{,}601$1 and $113{,}601$2 Hz; by contrast, the minimum reported COT changes only from 1.42 in the folded configuration to 1.32 in the erected configuration at the highest tested frequency (Huang et al., 2023).

3. Bluefin-21 and ambient-current estimation

Bluefin-21 denotes a specific unmanned underwater vehicle used to validate batch methods for estimating a steady, uniform flow-field from navigation data. The underlying planar kinematic model is

$113{,}601$3

with component form

$113{,}601$4

The estimation problem is to recover vehicle speed relative to water $113{,}601$5, current magnitude $113{,}601$6, and current direction $113{,}601$7 from noisy ground-velocity and heading measurements during a circular orbit or comparable large-heading-change maneuver (Wolek et al., 2024).

Three batch estimators are presented: a quadratic curve fit using $113{,}601$8, a nonlinear least-squares method using $113{,}601$9, and a nonlinear least-squares method using $100$0. Monte Carlo evaluation uses measurement-noise levels $100$1, heading intervals $100$2, 100 samples per dataset, and 250 randomized datasets per condition. The best-performing estimator is the nonlinear least-squares method using $100$3; for $100$4, it achieves vehicle-speed error $100$5 m/s, current-speed error $100$6 m/s, and current-direction error $100$7 under the assumed model (Wolek et al., 2024).

Field validation uses a Bluefin-21 operated by the U.S. Naval Research Laboratory, described as 20 ft in length and 21 inches in diameter, equipped with a fiber-optic-gyroscope inertial navigation system and a DVL navigation suite. On 2016-06-22, near Boston Harbor southeast of Nahant Bay, the vehicle executed circular orbits of 70 m and 120 m radius over more than five hours in a tide-driven flow. The resulting current estimates show good agreement with the ebb and flow directions reported by NOAA buoy BOS1132, located about 51 km away at Stellwagen Bank, while magnitude agreement is only described as following the general trend (Wolek et al., 2024).

4. Bluefin-Tuna and world-model robustness

Bluefin-Tuna appears as the organization releasing WISER, a robustness layer for world-model-based reinforcement learning. The central idea is to treat robustness as an online trust-calibration problem: observations that are too inconsistent with the learned dynamics should be rejected or selectively masked rather than passed directly to the policy. In the DreamerV3 setting, surprise is defined as the Bayesian surprise

$100$8

that is, the mismatch between the posterior latent inferred from the current observation and the predictive prior from the recurrent dynamics model (Zollicoffer et al., 30 Nov 2025).

Two rejection mechanisms are proposed. In multi-representation rejection sampling, the agent selects the subset of sensors or representations whose latent state minimizes surprise relative to the predictive prior; the appendix gives an $100$9 approximation that sorts sensors by surprise and iteratively masks the most suspicious ones. In single-representation rejection sampling, the agent either accepts the current observation or rejects it and switches into a predictive mode based on the world model’s internally predicted next state. In the DreamerV3-style experiments, the rejection threshold is set empirically as 5 standard deviations above the average reconstruction loss in the normal environment (Zollicoffer et al., 30 Nov 2025).

The method is evaluated in CARLA and Safety Gymnasium under Gaussian noise, occlusion, glare, jitter, chromatic aberration, and latency/lag, including settings with up to 5 sensor failures in CARLA. It is tested on two distinct world models, DreamerV3 and Cosmos Predict-2.5. For Cosmos, rejection sampling improves average PAI-Bench quality from 0.755 to 0.810, an absolute gain of 0.055 and a relative improvement of 7.11%, with especially large relative gains under Glare (11.98%) and Jitter (12.25%). The released codebase is WISER at https://github.com/Bluefin-Tuna/WISER (Zollicoffer et al., 30 Nov 2025).

5. BlueFin as a benchmark for financial spreadsheets

In exact title case, BlueFin most explicitly denotes the benchmark introduced in "BlueFin: Benchmarking LLM Agents on Financial Spreadsheets" (Kundurthy et al., 29 May 2026). It is designed for professional finance workflows and evaluates LLM agents on three categories of workbook-centered tasks: synthesis, manipulation, and interrogation or comprehension. The full dataset contains 131 tasks: 10 synthesis, 82 manipulation, and 39 interrogation. Across synthesis and manipulation, it defines 3,225 granular rubric criteria; the public release contains 11 tasks and 305 rubric criteria, while the main held-out evaluation uses 120 tasks divided into 75 manipulation, 9 synthesis, and 36 interrogation tasks (Kundurthy et al., 29 May 2026).

Task family	Count	Role
Synthesis	10	Build spreadsheet workbooks or major components from natural-language prompts
Manipulation	82	Modify existing finance workbooks while preserving logic and integration
Interrogation	39	Answer workbook-grounded finance questions

The benchmark is intentionally centered on dynamic workbook integrity rather than on isolated formula generation. Its six rubric dimensions are Formula Correctness, Model Integration, Output Validation, Perturbation, Presentation, and Pitfalls. The rubric inventory is highly granular: 1,082 Formula Correctness criteria, 765 Output Validation criteria, 468 Perturbation criteria, 383 Model Integration criteria, 334 Presentation criteria, and 193 Pitfalls criteria. Evaluation is carried out in a spreadsheet harness with 20 tools organized into six categories, including read, write, format, other, virtual, and recalculation functionality. Recalculation uses LibreOffice headless with iterative calculation enabled, and execute_python exposes the workbook as an in-memory openpyxl object inside a sandbox with no file I/O, no network, a restricted import allow-list, and a 30-second timeout (Kundurthy et al., 29 May 2026).

A central feature of BlueFin is its judge model. GPT-5.4, operating with reasoning_effort=high, serves as the agentic evaluator for generative tasks. In an inter-annotator alignment study over 384 binary criteria, the judge reaches parity with expert consensus at $129{,}794$0 and a macro-F1 score of 0.839. On the held-out set, no frontier model exceeds 50% overall: GPT-5.5 scores 49.6 overall, Claude Opus 4.7 scores 49.2, Gemini 3.1 Pro Preview scores 45.9, Claude Sonnet 4.6 scores 42.9, and Grok 4.20 scores 31.3. The paper emphasizes that models do relatively better on Formula Correctness, roughly 50–68%, than on Output Validation, about 20–48%, and especially Perturbation, about 15–37%, indicating a persistent gap between superficial plausibility and dynamic correctness (Kundurthy et al., 29 May 2026).

6. Bluefin tuna as biological subject and biomimetic template

The biological referent bluefin tuna enters the literature in two technically important ways: as an object of bioeconomic theory and as a source of biomimetic design. In the harvesting literature, the governing population and effort model is

$129{,}794$1

Near extinction, profitable harvesting is possible if

$129{,}794$2

The key concern is range contraction or hyperaggregation, which can keep local density high even as total abundance falls. For Atlantic bluefin tuna and Pacific bluefin tuna, the paper reports evidence of hyperaggregation, implying $129{,}794$3; it also cites an estimate $129{,}794$4 for wholesale fresh bluefin tuna in Japan. The resulting interpretation is that bluefin tuna may remain profitably harvestable at very low abundance without requiring dramatic rarity-price escalation (Burgess et al., 2014).

The same organism is also a direct biomimetic model. A tuna-inspired robotic fish simplifies its body shape from a real bluefin tuna using a NACA0018-derived profile and carries a silicone-cast morphing dorsal fin with four embedded rays and a four-bar linkage driven through magnetic coupling. The magnetic architecture is explicitly presented as a waterproofing solution for externally morphing appendages. Experimentally, the erected dorsal fin reduces peak-to-peak yaw across all tested cases, with improvements between 16.51% and 24.20%, but exerts minimal influence on speed and efficiency over the reported free-swimming operating range. This suggests that, in biomimetic design, the bluefin dorsal fin is most strongly supported as a stability-enhancing appendage rather than as a guaranteed source of cruising-speed or transport-economy gains (Huang et al., 2023).

Across these literatures, BlueFin therefore names not a single unified technology but a cluster of domain-specific entities and analogies. The strongest exact-title usage is the financial-spreadsheet benchmark, whereas underwater-robotics papers use the name chiefly as a design frame for fin-based propulsion, navigation, and sensing; Bluefin-Tuna uses it organizationally for robustness software; and the biological term bluefin tuna anchors both conservation arguments and biomimetic engineering (Kundurthy et al., 29 May 2026, Hamamatsu et al., 5 Feb 2025, Zollicoffer et al., 30 Nov 2025, Burgess et al., 2014).