Odysseus: A Multidisciplinary Research Label

• Odysseus is a polysemous research label that denotes distinct technical frameworks across astronomy, machine learning, decision theory, lunar science, and data systems.
• It integrates coordinated surveys, innovative AI methods, formal decision models, practical lunar lander experiments, and advanced data infrastructure applications.
• Its diverse applications highlight how disciplinary contexts fix the term’s meaning, ensuring self-consistent methodologies and actionable insights.

Odysseus is a polysemous designation in contemporary research. In current scholarly usage it denotes, depending on domain, a coordinated astronomical survey built around HST ULLYSES, several distinct machine-learning methods and frameworks, the Intuitive Machines Nova-C lunar lander used on the IM-1 mission, a classical decision-theoretic exemplar centered on the Sirens, and named systems in big-data infrastructure and urban monitoring (Espaillat et al., 2022, Luo et al., 11 Mar 2025, Hammond, 10 Jan 2026, Gopalswamy et al., 23 Mar 2026, Kim et al., 2014, Haycock et al., 2020). The term therefore functions less as a single referent than as a recurring research label whose meaning is fixed by disciplinary context.

1. Referential scope and naming

Across the cited literature, “Odysseus” is used in at least five technically distinct ways. In astronomy it denotes the ODYSSEUS survey, “Outflows and Disks around Young Stars: Synergies for the Exploration of ULLYSES Spectra,” a coordinated community program organized around the HST ULLYSES Director’s Discretionary program. In machine learning it names Dynamic Focus Decoding, a long-horizon VLM training framework, and a dual-steganography jailbreak paradigm. In decision theory it is the agent in “Odysseus and the Sirens Revisited,” used to motivate enlivened but truncated decision trees. In lunar science it is the Intuitive Machines Nova-C lander that carried the ROLSES instrument to the Moon. In data systems it appears both in Odysseus/DFS, an RDBMS-on-DFS architecture, and in Project Odysseus, a London mobility early-warning system (Pittman et al., 2022, Shi et al., 1 May 2026, Li et al., 23 Dec 2025, Hammond, 10 Jan 2026, Gopalswamy et al., 23 Mar 2026, Kim et al., 2014, Haycock et al., 2020).

Research domain	What “Odysseus” denotes	Distinguishing feature
Young-star astrophysics	ODYSSEUS survey	ULLYSES-centered accretion, outflow, and inner-disk program
Machine learning	Decoding method, RL framework, jailbreak paradigm	Adaptive decoding, 100+ turn VLM control, dual steganography
Decision theory	Odysseus–Sirens exemplar	Enlivened decision trees and subjective evaluations
Lunar space science	IM-1 Nova-C lander	ROLSES radio observations at the lunar surface
Data systems and analytics	Odysseus/DFS; Project Odysseus	DFS-backed RDBMS; network scan statistics for London mobility

A common misconception is to treat the term as if it retained a single stable meaning across papers. The literature shows the opposite. One ODYSSEUS astronomy study states that in that paper the term does not refer to the mythological hero or a spacecraft, whereas the ROLSES paper uses Odysseus precisely as the lander name; machine-learning papers use it as the name of an algorithm or framework; and the decision-theory paper returns to Homer’s tale as a formal motivating example (Pittman et al., 2022, Gopalswamy et al., 23 Mar 2026, Shi et al., 1 May 2026, Hammond, 10 Jan 2026).

2. ODYSSEUS in young-star and protoplanetary-disk research

In astrophysics, ODYSSEUS denotes a coordinated, multi-wavelength survey built around the HST ULLYSES program and aimed at accretion, outflows, disk irradiation, and star–disk coevolution in classical T Tauri stars (CTTS). The survey uses ULLYSES UV spectra together with contemporaneous optical, NIR, X-ray, and other facilities to measure how accretion flows depend on accretion rate and magnetic structures, determine where winds and jets are launched and how mass-loss rates compare with accretion, and establish the influence of FUV radiation on the chemistry of the warm inner regions of planet-forming disks. ODYSSEUS was explicitly designed to study the largest HST sample of T Tauri stars to date, about 60 CTTS in nine star-forming regions, spanning a range of ages and stellar masses (Espaillat et al., 2022, Pittman et al., 2022).

Methodologically, the survey combines physically motivated magnetospheric accretion shock models with inner-disk wall models. One Orion OB1b study applied multi-column shock modeling to HST/STIS G230L, G430L, and G750L data covering 1710–10,000 Å, together with the VLT/X-Shooter NIR arm covering 0.99–2.4 μm, and fit both the NUV accretion excess and the near-IR continuum excess from the frontally illuminated inner dust wall. In that nine-star sample, the derived mass accretion rates were $$0.5\text{–}17.2 \times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$$, with a median of $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$, and the wall temperatures and radii were $1200\text{–}1800$ K and $0.05\text{–}0.10$ AU, respectively. The same study identified extinction law choice as a central systematic, noting that the Whittet et al. (2004) law provides better fits to the UV continuum than Cardelli-like laws in Orion OB1b, with direct consequences for $A_V$, $L_{\rm acc}$, and $\dot M$ (Pittman et al., 2022).

Early ODYSSEUS work used individual systems as methodological demonstrations. The CVSO 109 paper established the survey’s core logic by combining HST UV spectra, optical and NIR spectroscopy, photometry, and X-rays to derive a self-consistent picture of accretion, outflow, and irradiation. For CVSO 109A, the best-fit accretion shock model yielded $$\dot{M}_{\rm acc} = 3.26^{+0.23}_{-0.24}\times 10^{-8}\,M_\odot\,\mathrm{yr^{-1}}$$, $L_{\rm acc} = 0.164^{+0.012}_{-0.012}\,L_\odot$, and an inner wall at $R_{\rm wall}\approx 0.09$ au for $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$0 K. The same system also showed FUV wind absorption and an unusual FUV irradiation field dominated by continuum rather than Ly$1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$1, illustrating how ODYSSEUS links accretion energetics to disk chemistry (Espaillat et al., 2022).

Subsequent monitoring papers extended the program to time-domain accretion physics. The UV paper in the four-target monitoring campaign found accretion variability ranging from short increases in accretion rate by up to a factor of 3 within 48 hours to longer decreases by a factor of 2.5 over the course of 1 year, and concluded that empirical relations between accretion rate and UV luminosity should not be used to estimate the accretion rate for an individual target. The photometric paper showed that all four benchmark CTTSs display significant variability, generally on days-long timescales and often due to periodicity associated with stellar rotation and stochastic accretion variability; it further concluded that photometry should be used with caution as a direct measure of accretion. The optical-spectra paper found magnetospheric truncation radii varying between $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$2 and reported that empirical line-luminosity relations reproduce HST-based accretion rates to within about 0.5 dex on average, but are less accurate for individual contemporaneous observations (Wendeborn et al., 2024, Wendeborn et al., 2024, Wendeborn et al., 2024).

Large-sample ODYSSEUS papers then generalized these results. A 67-star analysis of H$1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$3 flow modeling and UV–optical hotspot modeling found typical magnetospheric truncation radii to be almost half of the usually-assumed value of 5 stellar radii, showed that accretion-flow and shock-model accretion rates are consistent within $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$4 dex for contemporaneous observations, and demonstrated that up to 50% of the total accretion luminosity is at short wavelengths accessible only from space. A 47-star star–disk connection study measured corotation radii and fastness parameters, reported median $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$5 and median $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$6, concluded that most CTTS are in the spin-up regime, and argued that measured $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$7 values are consistent with observed ultra-short-period planet semi-major axes. A later Lupus study found the first evidence for parsec-scale spatial correlations of stellar magnetospheric inclinations, with correlation scales of about 3 pc, linking sub-au configurations to the geometry of natal filaments and expanding shells (Pittman et al., 1 Jul 2025, Pittman et al., 3 Sep 2025, Pittman et al., 30 Oct 2025).

Taken together, these papers define ODYSSEUS in astronomy as both an observing program and a modeling framework. Its unifying feature is not a single instrument or single observable, but a commitment to self-consistent accretion, extinction, outflow, and inner-disk inference across a homogeneous CTTS sample (Espaillat et al., 2022, Pittman et al., 1 Jul 2025).

3. Odysseus in machine learning and AI safety

In machine learning, “Odysseus” names at least three unrelated systems. The first is Dynamic Focus Decoding, introduced in “Odysseus Navigates the Sirens’ Song,” where the term designates a plug-and-play stochastic decoding wrapper for LLMs. DFD computes a token-level knowledge-awareness intensity from layer-wise KL divergences,

$1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$8

and maps it to a dynamic temperature $1.2\times 10^{-8}\,M_\odot\,\mathrm{yr}^{-1}$9, with the default best-performing transformation given as an exponential decay. The method is motivated by the claim that optimal decoding focus should change token-by-token: knowledge-aware steps require lower temperature, while stylistic steps can tolerate higher temperature. Across seven datasets, DFD improved both factuality and diversity, for example increasing TruthfulQA Top-k Truth+Info from 41.04 to 44.55 and StrategyQA nucleus accuracy from 65.40 to 68.60, while adding only about 1–7% FLOPs overhead depending on sequence length and model size (Luo et al., 11 Mar 2025).

A second computational use appears in “Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning.” There, Odysseus is an open training framework and recipe for long-horizon VLM control in Super Mario Land. The system uses RL-based training for 100+ turn decision-making, an adapted variant of PPO with a lightweight turn-level critic, positive-advantage filtering, and a small supervised initialization to supply strong action priors. The trained models achieved substantial gains across multiple levels of the game and at least 3 times average game progresses than frontier models, while also improving under in-game and cross-game generalization settings and maintaining general-domain capabilities (Shi et al., 1 May 2026).

A third use is explicitly adversarial. “Odysseus: Jailbreaking Commercial Multimodal LLM-integrated Systems via Dual Steganography” defines Odysseus as a jailbreak paradigm that hides malicious queries and responses inside benign-looking images. The attack uses dual steganography, function calling, error-correcting codes, and external extract and hide tools to move harmful content through channels that visible-content safety filters do not inspect. On benchmark datasets, the system achieved up to 99% attack success rate against several commercial multimodal systems and was presented as exposing a fundamental blind spot in defenses that assume malicious content must be explicitly visible in either the input or the output (Li et al., 23 Dec 2025).

These computational uses are methodologically unrelated and sometimes normatively opposed. One improves factual decoding, one stabilizes long-horizon embodied control, and one attacks safety systems. Their shared label is therefore nominal rather than theoretical (Luo et al., 11 Mar 2025, Shi et al., 1 May 2026, Li et al., 23 Dec 2025).

4. Odysseus and the Sirens in decision theory

In decision theory, Odysseus is used neither as an acronym nor as an engineered system, but as a formal motivating example. “Bounded Rationality with Subjective Evaluations in Enlivened but Truncated Decision Trees” revisits the Sirens episode in order to model decision problems whose future growth cannot be fully specified in advance. The paper presents a sequence of progressively richer trees—a naïve sailor’s model, a sophisticated sailor’s model, Kirke’s first model for Odysseus, and Kirke’s final model—and uses these to show how new contingencies and actions can enter the tree through an enlivening process (Hammond, 10 Jan 2026).

The conceptual move is from a fixed complete tree toward an enlivened decision tree whose structure, state space, and consequence domain can expand over time. A truncated tree $1200\text{–}1800$0 contains terminal nodes where full continuation subtrees are not modeled; at such points the agent assigns a real-valued subjective evaluation $1200\text{–}1800$1 rather than a fully specified consequence lottery. These scalar evaluations are then treated as terminal values within an extended Bayesian recursion. This allows Hammond to argue for an extended and refined form of Bayesian rationality even when the agent cannot model the full tree in detail (Hammond, 10 Jan 2026).

Within this framework, Odysseus’ binding to the mast is not treated as a case of dynamic inconsistency. The paper instead interprets it as a rational choice made in an enriched model of the initial decision problem, before reaching the Sirens, after Kirke’s advice has enlivened the tree to include wax, binding, and catastrophic downstream branches. This distinction matters because it relocates rationality from perfect foresight within a single complete model to coherent choice within a bounded and revisable model (Hammond, 10 Jan 2026).

The same paper extends the interpretation beyond Homeric narrative. Subjective evaluations at truncation nodes are explicitly linked to the kind of Monte Carlo tree search algorithm used by recent chess-playing software packages, and the broader framework is proposed as one way to rationalize the precautionary principle. Odysseus thus functions as a conceptual anchor for decision-making under model incompleteness rather than merely as a parable of self-control (Hammond, 10 Jan 2026).

5. Odysseus as a lunar lander and radio-science platform

In lunar science, Odysseus is Intuitive Machines’ Nova-C lander on the IM-1 mission, flown under NASA’s Commercial Lunar Payload Services program. The mission launched on 15 February 2024 and achieved the first U.S. lunar landing in over 50 years on 22 February 2024. The target was near Malapert A on the lunar near side close to the south pole, at about $1200\text{–}1800$2 S, and the lander became the platform for the ROLSES radio experiment (Gopalswamy et al., 23 Mar 2026).

ROLSES—the Radio-wave Observations at the Lunar Surface of the photoElectron Sheath instrument—was designed to characterize the radio and plasma-wave environment of the nearside lunar surface between 2 kHz and 30 MHz. On Odysseus, it consisted of four 2.5 meter radio monopole antennas, front-end electronics, digitization at 14 bits and 120 mega samples per second, and an FPGA performing onboard FFT-based spectral analysis. The frequency range was chosen specifically for phenomena below the terrestrial ionospheric cutoff, including radio waves from the Sun, the Milky Way galaxy, Jupiter, Earth’s auroral region, and ground-based radio transmitters (Gopalswamy et al., 23 Mar 2026).

The lander’s actual behavior was scientifically consequential. Odysseus tipped over after landing, altering antenna geometry, power, thermal conditions, and radio-frequency interference. One antenna deployed spontaneously during transit after overheating; another deployed inadvertently on the surface for the same reason. Even under these non-nominal conditions, transit observations detected terrestrial HF broadcast signals near the Moon, with a reported correlation coefficient of about 0.95 against simultaneous Radio JOVE observations from Florida (Gopalswamy et al., 23 Mar 2026).

The Odysseus mission is therefore significant in two senses. It served as an enabling host for lunar low-frequency radio observations, and it supplied an empirical measure of how a commercial lander behaves as a radio platform, including the extent to which lander-generated EMI constrains sensitivity. This, the paper suggests, directly informs the design of ROLSES 2 and future lunar radio arrays (Gopalswamy et al., 23 Mar 2026).

6. Odysseus in data infrastructure and urban monitoring

Two further technical uses of the name occur in information systems. “Odysseus/DFS: Integration of DBMS and Distributed File System for Transaction Processing of Big Data” defines Odysseus/DFS as an architecture that integrates the Odysseus relational DBMS with a distributed file system. The stated aim is to preserve SQL, schemas, indexes, and transactions while inheriting the scalability and reliability of DFS-based systems. The paper’s core storage abstraction is the meta DFS file, a logical DBMS file implemented as an ordered collection of DFS files, one per DFS block, together with a transaction-management method including recovery and concurrency control adapted to write-once-read-many DFS semantics (Kim et al., 2014).

The performance claims are operational rather than merely architectural. In transaction-processing experiments, Odysseus/DFS outperformed HBase, described there as a representative open-source NoSQL system, and compared with an RDBMS with local storage its performance was reported as comparable or marginally degraded. The design intent was explicit: to let a full DBMS utilize a virtually unlimited storage space provided by the DFS and thereby become suitable for big data analytics (Kim et al., 2014).

Project Odysseus, by contrast, is an urban analytics system for London. The COVID-19 paper describes it as a digital twin of busyness that ingests large-scale and heterogeneous mobility, transportation, and traffic datasets in order to support targeted interventions and policy-making. Its specific methodological contribution is an expectation-based scan statistic for networks, extending the Network Based Scan Statistic by making it expectation-based and by using stochastic processes for time-series forecasting. The system treats hourly geographically fixed time-series data on a road network, builds expected counts $1200\text{–}1800$3, and scans connected spatio-temporal regions for activity that is significantly higher or lower than expected, including a variant metric that focuses on identifying regions in which activity is quieter than expected (Haycock et al., 2020).

These two systems illustrate a recurring but nonexclusive pattern in the applied-computing uses of the name. In both cases, “Odysseus” denotes an infrastructure layer that mediates between a complex underlying substrate—DFS storage in one case, heterogeneous networked mobility data in the other—and a higher-level decision or query interface. This suggests a naming preference for systems that navigate between abstraction layers, though the papers themselves do not claim a common genealogy (Kim et al., 2014, Haycock et al., 2020).