Skip to main content
Back

Qwen-Flash Gets a Trust Score That Shuts It Down Mid-Optimization

A new gating mechanism stops LLMs from misleading Bayesian optimization when their confidence is wrong — tested on real molecule benchmarks

Paper, Full Access
  • Core idea: A new gating system dynamically decides how much to trust LLM suggestions per optimization objective, rather than applying a fixed trust weight.
  • Key mechanism: An objective-wise reputation-market tracks LLM accuracy separately for each property being optimized, updated online from experimental feedback.
  • Surprising finding: On ESOL, LLM confidence was positively correlated with prediction error — higher confidence meant worse suggestions.
  • Three-arm gate: The counterfactual gate chooses between using the LLM prior with confidence, without confidence, or abstaining entirely, per objective per step.
  • Useful negative result: Margin-based portfolio selection failed, showing that prior error alone is a bad proxy for optimization utility without acquisition-awareness.
  • Tested on: Three MoleculeNet benchmarks (ESOL, FreeSolv, Lipophilicity) using Qwen-Flash-generated priors; no public code released yet.

Using an LLM as a smart advisor during optimization sounds like a good idea , until it confidently steers you in the wrong direction. A new paper from Jiangyu Chen and colleagues tackles exactly this failure mode in the context of multi-objective Bayesian optimization (MOBO), proposing a dynamic gating system that decides, in real time, how much to trust an LLM's suggestions for each individual optimization objective.

The work sits at the intersection of two fast-moving areas: LLM-guided scientific search and Bayesian optimization for molecule discovery. The core insight is deceptively simple , an LLM might be a great advisor for one property of a molecule (say, drug-likeness) while being actively harmful for another (say, solubility). Treating it as uniformly trustworthy is a mistake.

The problem with blind trust

Bayesian optimization (BO) is the workhorse for expensive black-box optimization: you have a function you can only evaluate a limited number of times, so you build a probabilistic surrogate model and use it to pick the most promising candidates. Multi-objective Bayesian optimization (MOBO) extends this to handle multiple, often conflicting, expensive-to-evaluate objective functions simultaneously. The goal is to find the Pareto front , the set of solutions where you can't improve one objective without hurting another.

Lead optimization in drug discovery involves complex, multi-objective trade-offs, such as simultaneously improving solubility, permeability, and drug-likeness, under strict medicinal chemistry constraints. LLMs have emerged as a natural fit here: they encode vast chemical knowledge and can propose candidate molecules or predict which regions of chemical space are worth exploring. But there's a catch.

Modern LLMs , particularly those aligned by RLHF , can exhibit miscalibrated confidence, complicating uncertainty estimation without substantial validation data. In plain terms: when an LLM says it's 90% confident, that number often doesn't mean what you think it means. Verbalized confidence is systematically overconfident, with nominal 99% confidence intervals covering the true answer only 65% of the time on average. Plugging a miscalibrated prior directly into a BO loop can actively degrade optimization performance.