Mol-SFT: Supervised Fine-Tuning for Molecules

• Mol-SFT is the supervised fine-tuning of pre-trained molecular models using condition-molecule pairs to enable unified multimodal molecular tasks.
• The Molexar implementation uses a causal Transformer with Fragment-SELFIES, achieving 100% molecular validity and efficient multimodal generation.
• These methodologies integrate diverse chemical conditions into a single autoregressive framework while addressing challenges like data conflict and task interference.

Searching arXiv for papers defining or contextualizing Mol-SFT. Mol-SFT most commonly denotes molecular supervised fine-tuning: the adaptation of a pretrained model on molecular condition–output pairs so that a single model can follow chemistry-relevant instructions such as scalar property targets, pharmacophore constraints, protein-sequence context, or pocket-geometry context. In current usage, however, the term is not fully standardized. In "Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design" (Lin et al., 24 Jun 2026), Mol-SFT is the second-stage supervised training procedure that converts a pretrained autoregressive molecular decoder into a unified multimodal generator. In "Omni-Mol: Exploring Universal Convergent Space for Omni-Molecular Tasks" (2502.01074), the closely related paradigm is presented as unified instruction tuning across 15 molecular tasks. By contrast, "Socrates-Mol: Self-Oriented Cognitive Reasoning through Autonomous Trial-and-Error with Empirical-Bayesian Screening for Molecules" (Wang et al., 14 Nov 2025) positions itself as a different but complementary inference-time alternative to Mol-SFT, while "Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead" (Kang et al., 2 Oct 2025) does not define Mol-SFT as a named method and instead treats the relevant notion as an SFT-first methodology or SFT-based selection signal.

1. Terminological scope and definitional ambiguity

The literature in this set uses Mol-SFT in more than one sense. In the molecular generation setting, the term refers to supervised fine-tuning on molecular data or molecular conditions. In the reasoning-LLM post-training setting, the same shorthand is best understood as an SFT-stage methodology and, more specifically, as the post-SFT selection signal that the authors argue is often misleading (Kang et al., 2 Oct 2025).

Paper	Role of Mol-SFT	Key distinction
"Molexar" (Lin et al., 24 Jun 2026)	Supervised fine-tuning stage for a unified multimodal molecular generator	Parameter-updating, autoregressive molecular generation
"Omni-Mol" (2502.01074)	Mol-SFT-style unified instruction tuning	Emphasizes conflict control across heterogeneous tasks
"Socrates-Mol" (Wang et al., 14 Nov 2025)	Contrasted with Mol-SFT	Training-free or low-training inference paradigm
"Quagmires in SFT-RL Post-Training" (Kang et al., 2 Oct 2025)	Not a formal Mol-SFT algorithm	Critique of SFT-first model selection

A common misconception is that Mol-SFT names a single canonical algorithm. The available evidence does not support that reading. The papers instead support a broader definition: supervised fine-tuning methodologies that specialize general models for molecular tasks, with substantial variation in architecture, conditioning mechanism, and downstream objective. This also suggests that any encyclopedia treatment must distinguish between Mol-SFT as a molecular adaptation paradigm and SFT-stage scoring heuristics that may be called “Mol-SFT” only informally.

2. Molexar and the canonical multimodal Mol-SFT formulation

Molexar presents the clearest explicit realization of Mol-SFT as a molecular systems concept (Lin et al., 24 Jun 2026). The model is built on a Gemma2-style causal Transformer decoder with rotary positional embeddings, grouped-query attention, a mix of sliding-window and full-attention layers, and Gemma2 logit softcapping. The reported experimental configuration uses 16 transformer layers, hidden size 256, intermediate size 640, 4 attention heads, 1 key-value head, maximum sequence length 256, sliding-window size 128, and a 127-token vocabulary.

The training pipeline has two stages. First, a decoder is pretrained with causal language modeling on randomized Fragment-SELFIES strings. This stage teaches molecular grammar, fragment-tree traversal, and the distribution of valid molecules. Second, supervised fine-tuning (Mol-SFT) trains the same decoder on condition-molecule pairs. The conditions include scalar molecular properties, 2D pharmacophore fingerprints, protein-sequence embeddings, and pocket-geometry embeddings. The central design claim is that SFT is universal: all condition types are injected into a single sequence template so that the same decoder supports unconditional generation, property-controlled generation, pharmacophore-conditioned generation, target-aware generation from protein sequence, pocket-conditioned generation, and fragment continuation.

The shared prompt format is:

$p_i = 3 + 2i, \qquad i=0,\ldots,21,$4

This formulation matters because Molexar does not create separate architectures for each task. Instead, all modalities are routed through one shared autoregressive generation path. The paper explicitly frames this as a practical unified foundation for computational chemistry and drug-design workflows.

3. Conditioning mechanics and Fragment-SELFIES as the molecular substrate

Molexar’s Mol-SFT mechanism depends on two coupled design choices: Fragment-SELFIES as the molecular language and in-place replacement of value-token embeddings as the conditioning interface (Lin et al., 24 Jun 2026). Fragment-SELFIES represents a molecule as a BRICS fragment tree rather than as a raw atom string or corpus-specific fragment IDs. Its syntax includes [Frag], [Attach:M], [Frag@N], [pop], SELFIES-compatible fragment-body tokens, and [Dummy] attachment markers. For molecules not represented in BRICS-fragment form, the codec permits a fallback of the form [SELFIES]...[ENDSELFIES].

The representation is explicitly fragment-aware and validity-preserving. The paper attributes this to reconstruction of the fragment tree and reconnection of attachment sites, while SELFIES-compatible encoding inside fragments ensures arbitrary strings remain decodable into chemically valid structures. Molexar reports 100% validity in unconditional generation and states that this validity holds across downstream experiments.

Condition injection is implemented at fixed placeholder positions rather than by a separate encoder–decoder cross-attention stack. The placeholder position for the $i$th condition slot is

$p_i = 3 + 2i, \qquad i=0,\ldots,21,$

with the first twelve slots reserved for molecular properties, pharmacophore fingerprint, protein-sequence embedding, and pocket-geometry embedding. Raw condition values are converted into vectors by one-hot encoding for discrete scalars, radial basis function features for continuous scalars, or direct vector use for vector conditions; a two-layer projector maps each condition to the decoder hidden size $d=256$. The token embedding is then replaced in place:

$$\tilde{e}_t = \begin{cases} z_i, & t = p_i \text{ for an active condition } i,\ E(x_t), & \text{otherwise}. \end{cases}$$

This design has several consequences stated directly in the paper. The model does not add extra conditioning tokens, does not use cross-attention, and does not change the causal attention mask. All conditions therefore remain compatible with the same decoder and the same autoregressive key-value cache. A plausible implication is that Mol-SFT, in this formulation, is less a task-specific tuning recipe than a conditioning discipline for turning heterogeneous chemistry inputs into a single autoregressive continuation problem.

4. Generalist Mol-SFT and conflict-aware instruction tuning in Omni-Mol

Omni-Mol extends the Mol-SFT idea from a multimodal generator to a generalist AI chemist trained on heterogeneous molecular instructions (2502.01074). Its goal is to avoid what the authors call “conflict collapse”, where reaction tasks, property regression, captioning, and procedure generation interfere with one another during shared fine-tuning. The model therefore treats Mol-SFT not merely as data mixing, but as a convergence problem over representations, sample selection, and optimization.

The unified encoding mechanism casts each sample into Instruction $X_I$, Input $X_s$ as SELFIES strings, and Output $Y$, while also conditioning on graph-derived molecular features $H_G$. This allows reaction tasks, regression tasks, description tasks, and action generation to share one instruction-following interface. The reported task inventory comprises 15 tasks across 4 categories: forward reaction prediction, retrosynthesis, reagent prediction, solvent prediction, catalyst prediction, LogP prediction, SCF energy prediction, TPSA prediction, complexity prediction, molecular weight prediction, yield regression, HOMO/LUMO prediction, molecular captioning, description Q&A, and experimental procedure prediction.

A major departure from vanilla molecular SFT is the active-learning-driven data selection strategy. The paper states that this reduces the training dataset to 40% of its original volume while maintaining comparable performance to full-data training. It further reports that active selection with 40% of the data often beats random mixing and sometimes exceeds full-data training on certain tasks. The stated rationale is that molecular SFT is especially vulnerable to duplicated or overlapping samples, task imbalance, and compute spent on redundant examples.

Optimization is further stabilized by two additional modules: adaptive gradient stabilization and the anchor-and-reconcile MoE framework. The former modifies LoRA with an adaptive coefficient to control effective update magnitude; the latter expands later decoder layers into $N+1$ experts, with $N$ reconcile experts for specialized or conflict-resolving knowledge and 1 anchor expert for common shared knowledge. The paper presents these mechanisms as direct responses to three forms of negative conflict: representation conflicts, data-mixture conflicts, and optimization conflicts. This makes Omni-Mol a conflict-aware Mol-SFT system rather than a simple multitask fine-tuning baseline.

5. Empirical scope, capabilities, and reported performance

Molexar and Omni-Mol provide the principal empirical evidence for Mol-SFT as a scalable molecular modeling strategy (Lin et al., 24 Jun 2026, 2502.01074). Their reported capabilities are broad but differ in emphasis: Molexar concentrates on unified molecular generation under multimodal conditions, whereas Omni-Mol concentrates on universal instruction tuning across molecular prediction and generation tasks.

For Molexar, the reported SFT-related evaluations include single-property instruction following, multi-property instruction following, CrossDocked2020 target-conditioned generation, and MolGenBench. For single-property conditioning, the model is evaluated on all nine properties individually by generating 10,000 molecules per target, and the paper states that requested property modes align well with generated modes. For multi-property conditioning, the highlighted scenarios are HAC + RotBC, QED + SAS, and WT + LogP + HBDC, and the training schedule explicitly samples 129 property-only subsets: 9 singletons, 36 pairs, and 84 triples. On CrossDocked2020, three conditioning modes are reported: sequence-conditioned, pocket-conditioned, and pharmacophore-conditioned. The table values listed in the paper are Molexar (sequence): validity 1.00, QED 0.65, SA 0.82, Vina -7.25, high-affinity ratio 43.1; Molexar (pocket): validity 1.00, QED 0.65, SA 0.83, Vina -7.42, high-affinity ratio 53.0; and Molexar (pharmacophore): validity 1.00, QED 0.59, SA 0.71, Vina -6.79, high-affinity ratio 38.4. On MolGenBench, Molexar reports about 85% molecule pass rate and about 75% scaffold pass rate, exactly recovers the active-molecule SMILES for more than half of the targets in de novo generation, recovers nearly all active scaffolds, discovers the second-largest number of active molecules in hit-to-lead optimization, and achieves the highest potency among methods in that setting. The same paper also emphasizes efficiency, stating that Molexar has the lowest latency and lowest peak GPU memory across batch sizes and can sample up to 30,000 molecules in under 80 seconds on a single 80 GB GPU.

For Omni-Mol, the reported gains are framed against specialist baselines and prior molecular generalists. The model uses 1.7B parameters in the reported setup and achieves about 25% of the parameters of some specialist baselines. The paper reports roughly +1% over the prior SOTA generalist on forward reaction prediction, +4% on retrosynthesis, +13% on reagent prediction, +15% on molecular captioning, and +40% on description Q&A. Example table values include forward reaction prediction exact match 0.718 versus PRESTO 0.691 and InstructMol 0.536, retrosynthesis exact match 0.559 versus PRESTO 0.531 and InstructMol 0.407, and reagent prediction exact match 0.257 versus PRESTO 0.212 and InstructMol 0.129. Efficiency claims include up to 65% lower inference latency than PRESTO, 28% less memory, 1.78× higher token generation rate, and 41% lower GFLOPs.

Taken together, these results establish Mol-SFT as a practical route to either unified molecular generation or omni-molecular instruction following, depending on architectural choice. They also suggest that small or moderate parameter counts can remain competitive when the conditioning interface and training curriculum are carefully engineered.

6. Alternatives, complements, and critiques of SFT-centered methodology

Socrates-Mol is presented as a different but complementary paradigm rather than a substitute for Mol-SFT (Wang et al., 14 Nov 2025). It keeps model weights fixed and uses prompt/context engineering, retrieval, reflection, and self-consistency at inference time. The core mechanism is a reflective-prediction cycle in which the initial model output is treated as a prior, retrieved molecular cases and reflective notes provide evidence, and the refined output is treated as the posterior. For regression, the paper writes this as an update from $p_i = 3 + 2i, \qquad i=0,\ldots,21,$0 to $p_i = 3 + 2i, \qquad i=0,\ldots,21,$1. It introduces ranking tasks aligned with industrial screening priorities and uses cross-model self-consistency across five LLMs: Kimi, Qwen, DeepSeek, Doubao, and GLM. On amine solvent LogP prediction, the paper reports that regression achieves 72% MAE reduction and 112% R-squared improvement through self-consistency, with average MAE improving from 0.62502 to 0.2586 and average $p_i = 3 + 2i, \qquad i=0,\ldots,21,$2 from -4.06112 to 0.1591; ranking results improve from 43.6% basic average accuracy to 66.8% reflective average accuracy and 72% self-consistent accuracy. It also states that deployment cost is reduced by over 70% compared with full fine-tuning. The paper’s explicit conclusion is that Socrates-Mol is especially valuable when labeled data are sparse, deployment cost matters, iterative improvement through retrieval is desirable, and interpretable reflective reasoning is needed.

A separate critique arises from SFT-first post-training analysis outside the molecular domain (Kang et al., 2 Oct 2025). That paper challenges the assumption that high SFT scores translate to improved RLVR outcomes. Across hundreds of models up to 12B-parameter, trained with SFT and RLVR via GRPO and evaluated on 7 math benchmarks with up to 256 repetitions, it reports that high SFT scores can be biased toward simpler or more homogeneous data, repeated epochs, and data that is closer to the model’s existing behavior. It identifies generalization loss on held-out reasoning examples and Pass@large k, especially Pass@64, as better proxies for downstream RL performance, reporting improvements in $p_i = 3 + 2i, \qquad i=0,\ldots,21,$3 and Spearman rank correlation by up to 0.5 (2x) relative to direct prediction from pre-RL performance. This paper does not define Mol-SFT as a molecular method, but it provides a relevant caution: post-SFT benchmark accuracy alone is not necessarily a reliable model-selection criterion. A plausible implication is that Mol-SFT systems should not be evaluated solely by immediate post-fine-tuning gains when later-stage adaptation, search, or optimization is part of the workflow.

Within this literature, Mol-SFT therefore occupies a specific place. It is a parameter-updating molecular adaptation paradigm, often used to unify heterogeneous chemistry conditions or tasks within one model. Its principal strengths are shared molecular priors, unified interfaces, and multimodal or multitask compatibility. Its principal limitations, as suggested by adjacent work, concern task interference, data-mixture conflict, and the possibility that superficial post-SFT improvements may not reflect the most useful downstream behavior.