GenomeHarness, an agentic harness for adapting genome language models through controlled search over fine-tuning recipes, improves mean test MCC in 47 settings, and shows gains on Genomic Benchmarks and on tasks where the root recipe is unstable or poorly matched.
Abstract
Pretrained genome language models provide reusable representations for DNA sequence analysis, but turning them into reliable downstream predictors remains non-trivial. Their practical performance depends strongly on fine-tuning recipes, and default recipes reported in prior studies may be suboptimal for new tasks or model backbones, making weak downstream results difficult to interpret. These requirements place a substantial operational burden on many intended users, whose expertise is often centered on biological questions and interpretation rather than machine-learning engineering. Reliable use of genome language models therefore requires more than conventional AutoML-style tuning: it requires a systematic, budget-aware, and auditable procedure that lowers the barrier to downstream adaptation. We present GenomeHarness, an agentic harness for adapting genome language models through controlled search over fine-tuning recipes. GenomeHarness combines an AI agent for proposing and repairing recipe edits, a harness for protocol-constrained execution, resource management, and test isolation, and a Monte Carlo tree search controller for allocating search effort across recipe lineages. We evaluate GenomeHarness on DNABERT2 and NTv2-100M-Multi across the NT Benchmark and Genomic Benchmarks. Final evaluation is performed using three random seeds after recipe freezing. Across 52 model-task settings, GenomeHarness improves mean test MCC in 47 settings, including 24 of 26 DNABERT2 settings and 23 of 26 NTv2-100M-Multi settings. The gains are especially pronounced on Genomic Benchmarks and on tasks where the root recipe is unstable or poorly matched, such as human ocr ensembl task. Search traces further show that GenomeHarness progressively identifies stronger recipes, turning downstream adaptation into a controlled and auditable workflow rather than a manual tuning process.
Large Language Models (LLMs) have enabled capable agents across diverse applications. Beyond the foundation model, the performance of an agent is governed by the surrounding agent harness, including prompts, tools, control loops, etc. Automatically evolving this harness offers a promising pathway to agent improvement, yet existing approaches typically rely on greedy candidate selection and noisy self-generated feedback, rendering their gains susceptible to search collapse, task-specific overfitting, and poor verifiability. To tackle these challenges, we introduce HarnessBank, a trustworthy agent-harness self-evolution framework that pairs a task agent with a separate evolver agent for iterative failure diagnosis, harness generation, and evolution verification. HarnessBank maintains a Harness Gene Bank composed of high-performing harnesses of different semantic coordinates. Those harnesses are reinvented, recombined, screened, and selected during the self-evolution procedure. Moreover, we propose a Gated Harness Screening mechanism to efficiently filter high-quality harnesses and reduce the cost of evaluating numerous offspring harnesses. Across seven agent benchmarks, HarnessBank produces consistent performance improvements from 5.1% to 15.4%. Cross-model experiments further verify that the improvements come from the model-specific self-evolving process, instead of a universally optimal harness. Our code will be publicly available upon acceptance.
Xiaotian Luo, Dizhan Xue, Fengxingyu Wang et al.· 4 citations
Together, these results recast the evolved harness as a legible compensation layer, shaped jointly by the language's engineering demands and the model's behavioral gaps, rather than an opaque benchmark-tuned scaffold.
Siqi Yang, Qianlan Yang, Yu-Xiong Wang et al.· 2 citations
A representation-accessibility analysis of frozen genomic language models across regulatory, epigenetic, promoter, splice-site, and variant-effect prediction tasks shows that local biological signal is partially present in frozen representations, but is not always accessible through final pooled embeddings.
Nirjhor Datta, Swakkhar Shatabda, M. S. Rahman· 0 citations
Genomic Foundation Models (GFMs) are increasingly used for large-scale sequence analysis and generation. Compared with frontier language models, GFMs are typically smaller and frequently operate on long genomic sequences, with evaluation often requiring preservation of biologically meaningful structure and sequence-level relationships. Although low-precision post-training quantization (PTQ) has shown substantial memory and throughput benefits for general-purpose language models, it remains unclear whether these benefits transfer to GFMs given their distinct model scales, sequence characteristics, and evaluation requirements. We present an empirical case study of FP8 post-training quantization applied to GenomeOcean, a computationally efficient genomic foundation model with strong reported performance across diverse genomics tasks [Zhou et al., 2025]. Its range of model scales, from 100M to 4B parameters, provides a useful setting for examining how quantization effects vary with model size. We evaluate FP8 across two primary GFM inference regimes—embedding extraction and autoregressive generation—and assess its impact along two dimensions: biological fidelity relative to BF16 baselines and system-level efficiency in terms of throughput, memory usage, and energy efficiency. We find that FP8 largely preserves biological fidelity across the evaluated scales and inference regimes, while reducing GPU memory footprint at 4B scale and improving energy efficiency during autoregressive generation. However, realized throughput gains remain substantially below FP8’s theoretical 2× hardware ceiling, with a best-case improvement of 19.3% in autoregressive generation and benefits varying strongly by model scale and workload. Autoregressive generation shows the clearest gains, driven largely by KV-cache compression, whereas embedding extraction provides limited or negative throughput benefits at smaller model scales. We attribute this theory–practice gap to the interaction of model-scale effects, memory-system bottlenecks, and software-stack limitations. These findings highlight the need for workload-specific empirical evaluation before adopting low-precision inference in scientific foundation models. Code availability https://github.com/jgi-genomeocean/genomeocean_efficiency
Mutian Yu, Robert Egan, Fengchen Liu et al.· bioRxiv· 0 citations
GB.GeneUnet, an 837M-parameter transformer-based U-Net pretrained on 6 trillion tokens from multi-species genomes in OpenGenome2 is introduced, extending genomic context to 1 Mb with up to 100× inference speedup over GeneMoE, a preliminary MoE transformer baseline of similar model size pretrained on the same data.
Ning Sun, William de Vazelhes, Pan Li et al.· bioRxiv· 0 citations
