This paper presents DataSciBench, a comprehensive benchmark for evaluating Large Language Models (LLMs) in data science. Unlike existing benchmarks limited to single task, simple evaluation metrics, and readily available ground truth (GT), DataSciBench is built on curated, natural, and challenging prompts with complex evaluation criteria and uncertain GT. To bridge the gap, we develop a semi-automated GT generation pipeline, integrating LLM-based self-consistency and human verification to ensure accuracy, predefined task types, and aggregate functions (metrics). Furthermore, we introduce an innovative Intention-Function-Code (IFC) framework, assessing code execution outcomes through metrics and programmatic rules. Evaluating 26 models (8 API-based, 8 open-source general, 9 code generation, and 1 agentic models), our approach offers rigorous insights into LLM strengths and weaknesses. Experimental results show API-based models outperform open-source counterparts across all metrics, with DeepAnalyze-8B leading among open-sourced models. We release all code and data at https://github.com/THUDM/DataSciBench.

Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring **Guess–Refine–Perturb** dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce **Confident Decoding**, a training-free decoding strategy that dynamically selects a near-final layer through entropy-guided conservative backward search. We further formulate layer selection as an optimal stopping problem and show that, under explicit local-monotonicity and bounded-probe-noise conditions, the search rule filters terminal perturbations while bounding unfinished refinement. Experiments across dense and Mixture-of-Experts LLMs show that gains concentrate on challenging reasoning settings, including GPQA-Diamond, Omni-MATH, and HLE, while saturated tasks remain broadly stable, with no additional KV-cache memory and less than 2% latency increase. These results suggest dynamically bypassing final-layer perturbations can recover stronger reasoning behavior from aligned LLMs.