FaraGen: Synthetic Data for Molecules & CUAs

• FaraGen is a dual-purpose system that generates high-diversity synthetic data for both fragment-based molecule design and agentic computer use through staged pipelines.
• The f-RAG module enhances molecule optimization by dynamically retrieving hard and soft molecular fragments and applying genetic algorithms for evolutionary improvement.
• The CUA variant constructs and verifies multi-step web trajectories to produce large-scale, realistic datasets for training advanced agentic models.

FaraGen refers to two independently developed systems for high-throughput, high-diversity synthetic data generation, each targeting a different application domain: molecular design via fragment augmentation and agentic computer use via synthetic web trajectories. One variant, Fragment Retrieval-Augmented Generation (f-RAG), is a framework for fragment-based molecule generation that improves exploration and optimization by dynamic retrieval, injection, and evolution of molecular fragments (Lee et al., 2024). The other, the FaraGen engine for computer use agents (CUAs), is a scalable platform generating, solving, and verifying multi-step web-based tasks for training and evaluating agentic models such as Fara-7B (Awadallah et al., 24 Nov 2025). Both variants are characterized by staged pipelines, retrieval or proposal mechanisms, and iterative refinement with quantitative evaluation against baselines in their respective research fields.

1. Molecular Generation with Fragment Retrieval Augmentation

The f-RAG framework addresses limitations of conventional fragment-based molecule generation, which typically exhibits poor extrapolation beyond a fixed fragment vocabulary. Its architecture leverages a frozen, pre-trained molecular LLM (SAFE-GPT) and combines explicit and implicit fragment retrieval to maximize coverage of chemical space.

Architecture and Operation:

• SAFE-GPT Backbone: A Transformer-based autoregressive model over dot-concatenated “SAFE” molecular fragment strings, trained for fragment sequence completion.
• Hard Fragment Retrieval: Selects “hards,” which are strictly included building blocks (arms and/or linkers from a dynamically maintained vocabulary), and provides these as mandatory prefixes to SAFE-GPT. The selection alternates between linker-design and motif-extension tasks.
• Soft Fragment Retrieval and Injection: Retrieves K “softs” (informative but not explicitly used fragments) from the complementary vocabulary. These are embedded using lower SAFE-GPT layers and injected at the hidden state level via a trainable cross-attention Fragment Injection (FI) module, influencing the subsequent generation trajectory.

2. Fragment Scoring, Vocabulary Dynamics, and Genetic Modification

Fragment Scoring and Vocabulary Construction:

Fragments are ranked according to

$$\mathrm{score}(F_j) = \frac{1}{|S(F_j)|} \sum_{(x,y) \in S(F_j)} y, \qquad S(F_j) = \{(x, y) \mid F_j \subset x\}$$

where $y$ is the target property value for molecule $x$. Top fragments by score populate $\mathcal{V}_{\text{arm}}$ and $\mathcal{V}_{\text{linker}}$.

Fragment Injection and Self-Supervision:

The FI module fuses input embeddings with soft fragment representations using cross-attention, after which the upper SAFE-GPT layers generate candidate next fragments. The FI is optimized in a self-supervised regime: given input fragments, the system predicts missing/held-out fragments, scoring candidates using nearest-neighbor similarity (Tanimoto metric) and minimizing cross-entropy loss between predicted and ground truth fragments.

Vocabulary Refinement and Post-hoc Genetic Algorithms:

After each generation, the molecule is decomposed, fragments are scored and potentially incorporated into the vocabulary, supporting open-ended vocabulary growth. Post-hoc genetic modification operates over a population of top-performing molecules ($N_{\text{mol}}$), applying graph-based crossover (random bond reattachment) and mutational operators (including atom/bond edits and atom-type swapping). Offspring molecules are further scored, maintained in the population, and contribute new fragments to the vocabulary via a unified selection rule.

3. Exploration–Exploitation Trade-off and Empirical Performance in Molecular Generation

Exploratory Mechanisms:

• Soft fragment injection biases SAFE-GPT toward previously unexplored chemotypes.
• Vocabulary update allows the system to incorporate rare or novel fragments discovered during generation or via genetic algorithm (GA) modification.
• Post-hoc GA drives the search into underrepresented regions of chemical space.

Exploitation Mechanisms:

• Hard fragments from high-property vocabularies ensure optimization pressure toward desired biochemical properties.
• The frozen SAFE-GPT backbone maintains a strong generative prior.

Quantitative Results:

• PMO Benchmark: FaraGen attains higher diversity (Tanimoto dispersion 0.532), synthetic accessibility (SA score 2.026), and competitive novelty (0.800 above a 0.4 cutoff), outperforming alternatives like Genetic GFN and Mol GA in either sum AUC or key diversity metrics.
• Docking under multi-objective constraints: Demonstrates superior docking energy and hit rates versus VAE-, RL-, diffusion-, and retrieval-based methods.
• Ablation confirms both hard/soft retrieval and GA are essential: removing soft retrieval collapses diversity/novelty, while omitting GA limits exploration (Lee et al., 2024).

4. Synthetic Multi-Step Web Trajectory Generation for Computer Use Agents

FaraGen (for CUAs) is motivated by the scarcity of large, high-fidelity datasets capturing human-computer interaction sequences—particularly those involving pixel-level state and multi-turn action in live web environments. The system’s pipeline spans proposal, solution generation, and multi-level verification, enabling cost-effective large-scale data creation.

Pipeline Stages:

• Task Proposal: Uses three strategies: (A) Targeted-URL proposal via LLM intent generation on high-value sites (28% of tasks), (B) Agentic-URL exploration—agents invent tasks while actively browsing sampled sites (67% of tasks), and (C) Exemplar expansion—slot-filling over intent templates.
• Solution Generation: Employs a multi-agent system (built on Magentic-One), orchestrating high-level plans (Orchestrator), simulating browser actions (WebSurfer), and injecting user input at critical points (UserSimulator).
• Verification: Noisy trajectories from agents are filtered by an ensemble of LLM-based verifiers:
  • Alignment verifier checks text outputs against instructions.
  • Rubric verifier generates weighted sub-criteria, scoring satisfaction via

  $$\mathrm{Score} = \frac{\sum_i w_i\,\mathbb{1}[c_i\text{ met}]}{\sum_i w_i}$$ - Multimodal verifier analyzes screenshots to adjudicate successful completion. Trajectories passing any verifier are accepted as valid data points.

5. Data Statistics, Cost, and Impact for CUA Training

After filtering, FaraGen yields 145,603 verified trajectories (1,010,797 steps) across 70,117 unique web domains. The mean trajectory consists of 6.9 steps (range: 3–84), with a domain diversity measured at 0.48 (unique domains per trajectory). Table summaries reproduce the dataset and cost structure:

Statistic	Value	Notes
# Trajectories	145,603	verified
Total Steps	1,010,797	avg 6.9 per trajectory
Unique Domains	70,117	~0.48 new per trajectory

Component	o4-mini	o3	GPT-5
Orchestrator	\$0.32 | \$0.58	\$0.57
WebSurfer	\$0.25 | \$0.45	\$0.37
Alignment Verifier	\$0.00 | \$0.00	\$0.00
Rubric Verifier	\$y$00.03</td> <td>\$0.03
Multimodal Verifier	\$y$10.02</td> <td>\$0.02
Total per trajectory	\$y$21.08</td> <td>\$1.00

System throughput attains approximately 600 verified trajectories per hour on 40 GPU nodes, with yields varying by task structure (e.g., 3% for flight booking, up to 35% for shopping). The resultant FaraGen dataset enables training of competitive agentic CUA models such as Fara-7B, which outperforms comparably sized models (WebVoyager: 73.5%, Online-Mind2Web: 34.1%, WebTailBench: 38.4%) and approaches the performance of much larger frontier models (Awadallah et al., 24 Nov 2025).

6. Comparison with Prior Work and Prospective Extensions

In molecular generation, FaraGen (f-RAG) achieves a trade-off between exploitation (through curated fragment seeding and strong priors) and exploration (via dynamic vocabulary, soft retrieval, and GA). This balance yields best-in-class diversity, novelty, and synthesizability on established benchmarks, and enables discovery of non-trivial chemotypes and high-scoring drug candidates (Lee et al., 2024).

For CUAs, FaraGen provides orders-of-magnitude scale and real-world task variety not achieved by previous datasets, which were restricted to sandboxed or small-domain settings. The system’s three-tiered verification and multi-agent simulation allow it to cover dynamic content and multi-turn reasoning. Extensions under consideration include reinforcement-learning loops with user-in-the-loop corrections, coverage-driven adaptive proposal, and transfer to desktop UI, command-line, or mobile settings (Awadallah et al., 24 Nov 2025).

Conclusion:

FaraGen represents a methodological advance in both fragment-based molecule discovery and in agentic data synthesis for computer use. By orchestrating multi-stage retrieval, proposal, injection, and evolutionary modification, and through rigorous empirical evaluation, FaraGen systems deliver highly diverse, high-quality datasets that unlock state-of-the-art generative and agentic performance in their respective fields.