اكتب مراجعة على PREreview

Frontier large language models (LLMs) trained with extended reasoning, such as DeepSeek-R1 and OpenAI o1 , have established state-of-the-art benchmarks in formal logic, competitive programming, and mathematics. In interactive, empirical scientific sandboxes, however, their efficacy remains counter-intuitive. In the Bioinformatics Benchmark ( BixBench ), extended reasoning models paradoxically degraded in multi-turn environments, leading benchmark authors to exclude them from primary agentic evaluations. In this paper, we conduct the first systematic forensic audit of the internal cognitive monologue ( <think> … </think> ) of open-weight reasoning models ( DeepSeek-R1-Distill-Qwen-14B ) and matched instruction-tuned controls ( Qwen2.5-Coder-14B ) deployed inside containerized computational biology sandboxes on Nvidia H100 GPUs.

Our forensic analysis uncovers two distinct systemic failure modes: (1) The PubMed Prior Trap , in which reasoning models engage in epistemic rationalization within their cognitive stream, falsely convincing themselves that empirical datasets are missing and substituting empirical computation with speculative literature recall (e.g., estimating adjusted p -values based on “typical experimental outcomes”); and (2) Tag Language Nesting Collisions , in which models conflate agent-level XML action tags with Python syntax, producing terminal interpreter syntax errors. To resolve this paradox, we introduce Dual Track Decoupled Reasoning ( DTDR ), an agent scaffolding architecture that evicts internal cognitive monologues from multi-turn history, injects strict empirical grounding guardrails, and dynamically intercepts runtime tool invocations. On a stratified diagnostic benchmark spanning single-cell RNA-seq, bulk differential expression (DESeq2), and genomic variant prioritization ( n = 22 per arm), DTDR rescues the 14B reasoning model from 0.0% to 9.09% accuracy (2/22) , doubling the performance of the matched non-reasoning coder control ( 4.55%, 1/22 ) while slashing code execution errors by 81% (13.6% vs. 72.7%) and reducing average steps to solution by 66% (1.77 vs. 5.18 steps). All replication pipelines, prompts, and trajectory manifests are openly released.