DR-Venus: Venus Probe and Edge AI Agent

• DR-Venus is a dual-purpose platform, functioning as a deep-atmosphere probe for in situ Venus measurements and a 4B-parameter edge AI research agent for efficient deployment.
• The planetary probe employs advanced instruments like a quadrupole mass spectrometer and tunable laser spectrometer to deliver high-precision data on gases, isotopes, and aerosols during descent.
• The edge AI agent leverages a Qwen3-4B based architecture with supervised fine-tuning and reinforcement learning, achieving robust performance on multi-turn research benchmarks.

DR-Venus denotes both a flagship planetary atmospheric probe concept for Venus and, independently, a frontier small-parameter deep research agent for edge deployment. Both are prominent in their respective domains: planetary science and artificial intelligence. This entry provides an integrated reference for both senses, aligning as necessary with the sources (Garvin et al., 2020) and (Team et al., 21 Apr 2026).

1. DR-Venus as a Deep Atmosphere Probe Mission for Venus

DR-Venus refers to a proposed deep-atmosphere descent probe mission aimed at resolving key scientific questions about Venus’s evolution, composition, surface-atmosphere interactions, and potential habitability. Following the legacy of Venera-D, DAVINCI, and other precursor concepts, DR-Venus is optimized for in situ trace gas, isotope, and cloud property measurements across the lower atmosphere (70 km to surface) (Garvin et al., 2020).

1.1 Scientific Objectives and Data Products

DR-Venus is structured around three major scientific objectives:

• Composition and Evolution: Determining bulk and isotopic composition (target ratios: D/H, $^{13}$C/$^{12}$C, $^{15}$N/$^{14}$N, noble gases including $^{3}$He/$^{4}$He, $^{36}$Ar/$^{38}$Ar, Xe isotopes) in the lower atmosphere, constraining Venus’s origin, history of volatile inventory, and planetary differentiation processes.
• Volcanic Outgassing and Atmospheric Chemistry: Quantifying trace gases (SO$_2$, CO, OCS, HCl, HF, CO$_2$ at supercritical conditions), detecting signatures of volcanic activity or thermochemical disequilibrium, and elucidating sulfur chemistry and weathering reactions at the surface.
• Potential Habitability and Biosignatures: Investigating the $^{12}$050–60 km cloud deck for evidence of chemical disequilibria or biomarkers (PH$^{12}$1, NH$^{12}$2, organics), assessing concentrations and isotopic features of possible biogenic gases.

Expected data products include high-precision vertical profiles of major gases ($^{12}$32%), noble gases ($^{12}$45%), key isotope ratios ($^{12}$50.5–1$^{12}$6), trace species (detection to 1 ppbv), aerosol microphysics ($^{12}$710%), and descent imaging at NIR windows down to 5 m resolution (Garvin et al., 2020).

2. Mission Architecture: Timeline, Engineering, and Environmental Models

The DR-Venus mission profile consists of planetary transfer, probe release, atmospheric entry, parachute-mediated descent, and end-of-mission data relay:

• Trajectory and Entry: Post-Earth launch, a $^{12}$8120–150 day transfer to Venus culminates in probe release $^{12}$90.5 AU from Venus. Atmospheric entry commences at 125 km with $^{15}$011 km/s velocity and peak heating near 2000 K for 120 s.
• Descent Dynamics: Parachute deployment phases dictate descent rates ($^{15}$125 m/s to 10 m/s from 50 km to surface). The descent lasts 85 minutes, with up to 1 h of survival at surface $^{15}$2 K, $^{15}$3 bar.
• Analytical Payload: Instruments include a quadrupole mass spectrometer (QMS, 1–150 amu), tunable laser spectrometer (TLS), gas chromatograph (GC), nephelometer, environmental sensors (wide $^{15}$4,$^{15}$5 dynamic range), and NIR descent imaging microprobe. Sensing capabilities extend to $^{15}$61 ppbv sensitivity for targeted trace species and $^{15}$7–$^{15}$8 for isotope measurements.

Descent equations model pressure as $^{15}$9 (with $^{14}$0 bar, $^{14}$1 km), and descend velocity as $^{14}$2. Thermal design safeguards instrument temperatures $^{14}$3450 K via insulation and phase-change materials (Garvin et al., 2020).

3. Data Analysis and Science Return

Data products from DR-Venus enable:

• Discrimination of Photochemical versus Volcanic Processes: High vertical resolution of trace species and isotopes allows classification of atmospheric disequilibrium and local surface–atmosphere exchanges.
• Hydrodynamical and Climate Model Inputs: D/H ratio versus altitude, noble gas patterns, and aerosol profiles supply critical constraints for historical water loss, climate runaway, and exoplanet analog inference.
• Volcanism and Habitability: SO$^{14}$4 and CO spikes support active volcanism hypotheses, while PH$^{14}$5/NH$^{14}$6 quantification in clouds sets upper limits for potential biosignature fluxes.
• Surface-Atmosphere Linkages: Imaging at descent and near-surface provides terrain typing and supports correlation of atmospheric and lithospheric composition.

4. DR-Venus as an Edge-Scale Deep Research Agent

Independent of the planetary probe context, DR-Venus also names a 4B-parameter deep research agent for edge deployment, trained exclusively on open-data with specialized techniques for small-footprint efficacy (Team et al., 21 Apr 2026).

4.1 Model Backbone and Deployment

DR-Venus utilizes the Qwen3-4B-Thinking-2507 architecture—a causal LLM with $^{14}$74B parameters. This scale permits deployment on edge servers and devices, offering cost and latency advantages, and privacy via local data handling.

4.2 Data Pipeline and Training Protocol

DR-Venus’s agentic supervised fine-tuning (SFT) stage uses a REDSearcher open-data corpus (originally 10,001 trajectories), applying:

• Environment and Format Normalization
• Unsupported Tool/Redundancy Removal
• Bilingual Judge Correctness Filtering (retaining only high-confidence answers)
• Turn-Aware Resampling emphasizing long-horizon (T$^{14}$8100) trajectories

Final SFT training applies next-token log-probability loss over model outputs (actions, reasoning), masking environmental observations.

4.3 Reinforcement Learning and IGPO

To maximize reliability in multi-turn research, DR-Venus applies agentic RL via Information Gain-based Policy Optimization (IGPO):

• Turn-Level Reward: Immediate IG reward at each turn $^{14}$9
• Format Penalty: Turnwise negative rewards for output malformation
• Reward Normalization, IG Scaling, and Discounted Credit Assignment
• Objective Function: Clipped GRPO-style per-token policy update with per-token advantages; see equation (10):

$^{3}$0

By distributing reward at every turn and token, IGPO achieves dense supervision and robust long-horizon credit assignment (Team et al., 21 Apr 2026).

5. Benchmark Performance and Comparative Analysis

On standard agentic deep-research benchmarks (BrowseComp, BrowseComp-ZH, xBench-DS-2505, xBench-DS-2510, DeepSearchQA), DR-Venus-4B demonstrates:

Model	BrowseComp	xBench-DS-2505	DeepSearchQA
AgentCPM-4B-Explore	24.1	70.0	32.8
DR-Venus-4B-SFT	26.8	69.0	37.7
DR-Venus-4B-RL (final)	29.1	74.7	39.6
Tongyi-DR-30B	43.4	75.0	–

DR-Venus-4B matches or outperforms previous $^{3}$19B open agents and narrows the gap to 30B-class systems, with RL boosting Pass@1 by several points across tasks. Analysis reveals a "capability ceiling" (Pass@16 $^{3}$278.5% on BrowseComp-ZH), and emphasizes the importance of browsing action for trajectory accuracy (Team et al., 21 Apr 2026).

6. Broader Scientific and Technological Implications

6.1 Venus Mission

• Planetary Science: DR-Venus delivers boundary conditions essential for global climate modeling and terrestrial planet comparative analysis, supporting resolution of early water inventory, greenhouse onset, and crust-atmosphere interaction (Garvin et al., 2020).
• Astrobiology and Exoplanets: In situ constraints from DR-Venus inform radiative-convective models for JWST-era exoplanet studies, refine biosignature detection criteria, and provide analogue data for hot, CO$^{3}$3-dominated atmospheres.

6.2 Edge-Scale Agent Deployment

• Practical AI: DR-Venus demonstrates the viability of small, open-data-trained agents for sophisticated multi-turn research, highlighting the latent capability and deployment efficiencies available at $^{3}$44B parameters (Team et al., 21 Apr 2026).
• Reproducibility: Open release of models, code, and data pipelines facilitates further research and benchmarking.

7. References and Data Availability

The DR-Venus research agent models, codebase, and data construction pipelines are available via https://github.com/inclusionAI/DR-Venus and HuggingFace. The planetary mission details synthesize findings and design elements from (Garvin et al., 2020), complemented by historical comparative missions as outlined in (Wilson, 2017).

---

In both planetary science and edge AI, DR-Venus designates projects defined by a meticulous approach to data quality, robust engineering, and detailed benchmarking, advancing the frontiers of Venus exploration and research-centric AI deployment.