Pharos: Disambiguation Across Research Domains

• Pharos is a multi-faceted research term defining distinct platforms across domains such as biomedical imaging, reinforcement learning, and cultural heritage.
• It underpins fairness-aware CT benchmarks, NIH drug target classification, and ESG report parsing with measurable performance metrics.
• Applications extend to real-time accelerator design, UAV airspace management, live robot programming, and laser instrumentation, highlighting robust evaluation practices.

Pharos is a recurrent system name in contemporary research, used for multiple independent platforms, benchmarks, infrastructures, and instruments rather than a single canonical artifact. Recent literature applies the name to a fairness-aware chest CT challenge ecosystem, the NIH Illuminating the Druggable Genome portal, an open-source reinforcement-learning benchmarking library, a cultural-heritage consortium and research platform, a smartwatch navigation system, a multimodal ESG-report parser, a multi-UAV airspace manager, a real-time accelerator design framework, and a femtosecond laser platform (Parikh et al., 13 Mar 2026, Wang et al., 2024, Conserva et al., 21 Sep 2025, Daquino et al., 21 Jan 2026, Wenig et al., 2019, Chen et al., 20 Nov 2025, Sun et al., 6 Jul 2026, Ji et al., 7 Apr 2026, Ejopu et al., 1 Sep 2025).

1. Nomenclature and disambiguation

A common source of confusion is that Pharos, PHAROS, pharos, and PhaROS do not denote one research lineage. The literature uses the name across unrelated domains, and capitalization often tracks distinct projects rather than stylistic variation.

Referent	Domain	Defining role
PHAROS-AIF-MIH / Fair Disease Diagnosis Challenge (Parikh et al., 13 Mar 2026)	Medical imaging	Fairness-aware chest CT benchmark
Pharos portal of NIH IDG (Wang et al., 2024)	Biomedical informatics	Catalog of human proteins and target taxonomy
pharos library (Conserva et al., 21 Sep 2025)	Deep RL	Representation-aware benchmarking library
Pharos-ESG (Chen et al., 20 Nov 2025)	Document AI	ESG report parsing and labeling framework
PHAROS consortium / artresearch.net (Daquino et al., 21 Jan 2026)	Cultural heritage	Federated art historical photo archives
Pharos smartwatch system (Wenig et al., 2019)	Navigation HCI	Global-landmark pedestrian guidance
Pharos airspace system (Sun et al., 6 Jul 2026)	UAV coordination	Collaborative multi-UAV airspace management
PHAROS accelerator framework (Ji et al., 7 Apr 2026)	Real-time systems	SRT-oriented HA design framework
Light Conversion PHAROS (Ejopu et al., 1 Sep 2025)	Laser instrumentation	Femtosecond source for TPA characterization
PhaROS (Estefó et al., 2014)	Robotics	Pharo–ROS bridge for live programming

The name therefore functions as a homonym in research communication. Context is decisive: in biomedical work, “Pharos” usually refers either to the NIH IDG portal or to PHAROS-branded medical-imaging challenges; in systems work it can denote accelerator DSE or UAV airspace management; in robotics, the related but distinct spelling PhaROS refers to a Pharo–ROS bridge (Estefó et al., 2014).

2. Biomedical and clinical meanings

In biomedical informatics, Pharos is the public portal of the NIH Illuminating the Druggable Genome program. Its central organizing device is the target taxonomy Tclin, Tchem, Tbio, and Tdark, which classifies proteins by the depth and type of supporting evidence. The portal aggregates disease links, drug interactions, pathways, and literature, and is used as the source of label sets in downstream computational studies such as GAN-TAT, which maps Pharos Tclin labels onto a 6,048-node signaling PIN with 20,697 directed edges and 324 retained extended features after filtering (Wang et al., 2024).

A separate clinical meaning appears in the PHAROS-AIF-MIH and PHAROS Fair Disease Diagnosis chest CT challenges. In one fairness-aware task, scan-level classification is performed over four classes—Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma—and evaluation explicitly averages per-gender macro-F1, so improvement for one gender cannot compensate for degraded performance for the other. The challenge score is defined as

$$\text{Score}=\frac{\text{MacroF1}^{(\text{male})}+\text{MacroF1}^{(\text{female})}}{2}.$$

The associated study reports 889 CT scans total, with 734 training and 155 validation, and emphasizes the extreme scarcity of the Female SCC subgroup, which contains 18 female scans versus 91 male scans (Parikh et al., 13 Mar 2026).

The PHAROS-AIF-MIH benchmark is also used as a multi-source robustness testbed. One study evaluates binary COVID-19 detection and four-class disease categorization under scanner, protocol, and demographic shift, using both 2.5D and 3D branches. On the validation set, the ensemble achieves 94.48% accuracy and 0.9426 Macro F1-score for the binary task, while the 2.5D DINOv3 branch achieves 79.35% accuracy and 0.7497 Macro F1-score for the multi-class task (Yang et al., 16 Mar 2026).

The challenge literature also illustrates reporting nuances. In the fairness-aware MIL paper, the abstract states that the best single fold reaches 0.759, whereas the per-fold table reports 0.727 as the highest value before threshold optimization. The same work reports a mean validation competition score of 0.685 with standard deviation ± 0.030, and female macro-F1 0.691 ± 0.030 versus male macro-F1 0.679 ± 0.068, which the authors interpret as evidence of a reduced fairness gap (Parikh et al., 13 Mar 2026).

3. Machine-learning and evaluation frameworks

In deep RL, pharos is an open-source benchmarking library introduced to make principled evaluation practical in the non-tabular regime. Its distinguishing property is explicit control over both the underlying MDP and the observation function, allowing the same environment to be exposed as a normalized state vector or as one or more pixel-based renderings. This design supports computation of tabular hardness measures such as diameter

$$D := \sup\limits_{s' \in \mathcal{S}} \inf_\pi T^\pi_{s \to s'}$$

and suboptimality gaps

$\Delta(s,a) := V^*(s) - Q^*(s,a),$

while separately varying representation. The central empirical result is that representation hardness dominates non-tabular difficulty: the single mixed regression model has $R^2 = 0.09$, the image-only model $R^2 = 0.3$, and the vector-based model $R^2 = 0.6$, with environment-specific fits ranging from $R^2 = 0.42$ in simple_grid to $R^2 = 0.96$ in frozen_lake (Conserva et al., 21 Sep 2025).

A different ML use appears in adversarial hashing. There, pharos is not a library or benchmark but a semantics-preserving binary target code used in the Pharos-guided Attack (PgA). The code is computed in closed form as

$$b^\star = \operatorname{sign}\!\left(\sum_{i=1}^{N_p} w_i b_i^{(p)} - \sum_{j=1}^{N_n} w_j b_j^{(n)}\right),$$

and the attack maximizes the Hamming distance between the adversarial query’s hash and this code. On DPH with 32-bit codes, PgA reduces MAP to 15.18 on FLICKR-25K, 11.53 on NUS-WIDE, and 9.41 on MS-COCO, with 0.04s per image, outperforming HAG, SDHA, DHTA, THA, and ProS-GAN in both attack strength and speed (Wang et al., 2023).

These two uses share a concern with principled evaluation but are otherwise unrelated. One studies benchmark construction under representation shift; the other defines a target code for white-box adversarial optimization.

4. Knowledge infrastructures, document parsing, and uncertainty

In cultural heritage, PHAROS is an association of 13 art historical photo archives in Europe and North America that collaboratively publish and reconcile metadata for cross-collection discovery. Collectively, the holdings comprise approximately 31 million photographic documents, while the pilot on artresearch.net currently integrates data from seven institutions amounting to about 1.7 million photographs of roughly 1 million artworks. The platform is built on ResearchSpace and uses a CIDOC CRM application profile to harmonize heterogeneous archival descriptions while preserving provenance and institutional perspective (Daquino et al., 21 Jan 2026).

A central conceptual contribution of this PHAROS literature is that reconciliation is “seldom a one-to-one operation.” The system therefore uses rdfs:seeAlso rather than owl:sameAs when equivalence is possible but not certain, creates anonymous or collective entities such as “Bellini,” groups unstable cases under umbrella terms such as “Böhm,” and preserves provenance through named graphs and UI affordances. Cross-authority harmonization reveals that 27% of cross-reconciled URIs show inconsistencies across authorities, which is treated not as an anomaly but as a structural consequence of the Open World Assumption and heterogeneous cataloguing traditions (Daquino et al., 21 Jan 2026).

In document AI, Pharos-ESG is a unified framework for ESG reports that combines reading-order modeling, ToC-guided hierarchy reconstruction, multimodal narration, and hierarchical ESG/GRI/sentiment labeling. Its outputs include block-wise reading order, heading paths, multimodal narrative text, and labels. The accompanying Aurora-ESG dataset contains 24,409 reports and over 8 million content blocks across Mainland China, Hong Kong, and the U.S. On the comprehensive ESG report analysis benchmark, Pharos-ESG reports precision 92.23, recall 95.00, F1 93.59, and ROKT 0.92; for ToC-body alignment it reports TBTA 92.46; and for hierarchical labeling macro-F1 86.32 with HLA 94.78 (Chen et al., 20 Nov 2025).

Taken together, these infrastructures show two distinct senses of “Pharos”: one centered on provenance-aware aggregation under uncertainty, the other on structure recovery and multimodal normalization. Both treat heterogeneity as a primary design constraint rather than a residual nuisance.

5. Guidance and coordination in physical environments

In navigation HCI, Pharos is a smartwatch-based pedestrian navigation system that augments turn-by-turn instructions with a single global landmark rather than multiple local landmarks. Visibility is inferred from Google Street View imagery using landmark-specific detectors; across three landmarks, CNN + sliding window achieves F1 = 84.13, outperforming CNN + selective search and classical descriptor-plus-SVM variants. In a within-subjects study in Bremen with $N = 12$, Pharos raises mean confidence from 6.1 to 6.9, reduces mean looks at the smartwatch from 27.6 to 19.6, and improves cognitive-map correlation from 0.51 to 0.63, while time and error counts remain statistically unchanged (Wenig et al., 2019).

In urban UAV coordination, Pharos denotes a collaborative multi-UAV airspace management system that allocates non-overlapping exclusive cuboid spaces instead of issuing fine-grained trajectory commands. It is trained with MAPPO, uses a 24-dimensional local observation including repulsive UAV features and seven fear predictions, and optimizes a shared reward balancing collision penalty, human fear penalty, and progress:

$$D := \sup\limits_{s' \in \mathcal{S}} \inf_\pi T^\pi_{s \to s'}$$0

The paper reports that Pharos improves spatial entropy by 70.82% versus Ipopt and 2.03% versus A-star. It also reports a 52.72% reduction in average human fear, but the benchmark comparison is internally inconsistent: the abstract states this reduction is relative to Ipopt, whereas the detailed results section attributes the 52.72% reduction to comparison with A, and the tabulated fear values show **Ipopt* lower than Pharos because Ipopt favors hovering and sacrifices long-term progress (Sun et al., 6 Jul 2026).

These two systems are methodologically unrelated, but both use the name for guidance under uncertainty: one orients pedestrians through persistent visual anchors, the other coordinates aircraft through coarse-grained spatial exclusivity and human-centric reward shaping.

6. Robotics, instrumentation, and real-time systems

The spelling PhaROS designates a bridge between ROS and the dynamic language Pharo, not a member of the broader “Pharos” naming family in any substantive technical sense. PhaROS reifies ROS nodes as objects, exposes topics and parameters through a Pharo facade, supports restarting nodes and changing parameters at run time, and is paired with Live Robot Programming (LRP) to satisfy four requirements for live ROS programming: starting and stopping nodes repeatedly, changing node parameters at run time, partially or fully replacing node code at run time, and changing node connections at run time (Estefó et al., 2014).

In instrumentation, Light Conversion PHAROS is a femtosecond laser platform used with the ORPHEUS-HP OPA for two-photon absorption characterization of solid-state sensors. In the reported setup the PHAROS pump provides 260 fs pulses, 200 μJ pulse energy, and repetition rate up to 100 kHz, while the OPA covers 310–16000 nm. At 1550 nm and 405 nm, 150 fs pulses are available. Knife-edge characterization in silicon yields a Rayleigh length of $$D := \sup\limits_{s' \in \mathcal{S}} \inf_\pi T^\pi_{s \to s'}$$1 and a beam waist radius of $$D := \sup\limits_{s' \in \mathcal{S}} \inf_\pi T^\pi_{s \to s'}$$2, and the TPA response shows the expected quadratic dependence on pulse energy (Ejopu et al., 1 Sep 2025).

In real-time computer architecture, PHAROS is a framework for pipelined heterogeneous accelerators under soft real-time constraints. It supports FIFO and EDF scheduling, introduces tile-granularity preemption, and embeds schedulability into design-space exploration through the per-accelerator utilization condition

$$D := \sup\limits_{s' \in \mathcal{S}} \inf_\pi T^\pi_{s \to s'}$$3

Its DSE objective is to minimize the maximum accelerator utilization, and the evaluation reports 1.44×–2.28× more feasible solutions than the best throughput-guided baseline, with one case study yielding 49 feasible task sets versus 13 for the throughput-oriented alternative. With beam width $$D := \sup\limits_{s' \in \mathcal{S}} \inf_\pi T^\pi_{s \to s'}$$4, the method finds the first feasible solution 13.3× faster and completes 117.2× faster than brute-force search (Ji et al., 7 Apr 2026).

Across these uses, “Pharos” functions as a name for systems that impose structure on dynamic execution, whether the substrate is a live robot program, a tunable femtosecond optical beam, or a deadline-constrained accelerator pipeline. This suggests that the principal encyclopedic challenge is not thematic unification but accurate disambiguation across domains.