What We Observe as LLM Behavior Can Be a Side-effect of Inference Backend

Progress in LLMs is increasingly measured through standardized benchmarks, where state-of-the-art improvements are often separated by fractions of a percentage point. At the same time, the computational cost of evaluating modern LLMs has driven widespread adoption of specialized inference backends, software systems that execute trained models efficiently at inference time. While critical for scalability, system-level optimizations, such as custom CUDA kernels and reduced-precision arithmetic, can alter token probabilities and introduce non-determinism, possibly cascading into divergent generation. In this work, we first survey the inference landscape, identifying 200 distinct engines, and analyze 35,000 ML publications, finding that the specific inference stack is rarely reported despite this widespread diversity. We then present a systematic empirical study of how inference backends affect LLM benchmark results. Holding model weights, decoding parameters, and hardware constant, we evaluate five widely used inference engines, including vLLM, SGLang, and llama$.$cpp, across multiple open-weight models and established benchmarks. We show that the choice of backend alone can shift benchmark scores by up to 16.6 percentage points and induce high rates of output disagreement. By isolating backend optimizations and tracing the execution pipeline, we find this divergence is driven by system-level optimizations like prefix caching and CUDA graphs, custom kernels, and engine-specific defaults in logit processing. Our findings identify the inference backend as a previously unreported but consequential hyperparameter in the evaluation of LLM and advocate standardized reporting of inference stacks to improve the reproducibility and interpretability of benchmark comparisons.

AI evaluations are shifting toward harder tasks that benefit from longer trajectories involving tool use and iterative problem solving. As a result, performance is increasingly sensitive to the amount and allocation of compute available at test time ("inference compute"). Yet many evaluations still report performance at a single restrictive budget, meaning that low scores may reflect the evaluation setup rather than the model's underlying capability. To test this, we evaluate up to 12 frontier language models on seven challenging benchmarks spanning software engineering, mathematics, medicine, and cybersecurity. We use a controlled setup combining three simple inference-scaling interventions: larger token budgets, context compaction, and repeated submission attempts, guided either by the model itself or by minimal correctness feedback. We find three main results. First, larger token budgets substantially improve performance on benchmarks across multiple domains, including cybersecurity, FrontierMath, Humanity's Last Exam, and TerminalBench. Second, fixed-budget evaluations can increasingly understate frontier capability as models advance. Newer models reach higher performance at large budgets, where they unlock harder tasks and solve them more reliably. Third, benchmarks differ in which inference-scaling methods help most: repeated submission broadly improves performance, but the value of larger token budgets, external feedback, and parallel attempts varies by benchmark. Overall, our results show that benchmark scores are protocol-dependent. We therefore argue that evaluations should report capability as a function of inference-time compute, specify protocol choices explicitly, and compare model generations over a large shared compute range at matched budgets, especially in safety- or policy-relevant settings.

LLM inference optimizations report speedups on different models, GPUs, prompts, and quality metrics, making them hard to compare or combine. We build a cost, quality, and latency Pareto atlas to identify the best configurations for different deployment constraints. Since exhaustive testing is impractical, we measure 54 configurations of Qwen2.5-7B-Instruct running on vLLM 0.12 across L4, A100, and H100 GPUs and use these anchors to calibrate a simulator. It reproduces measurements at anchored batch sizes, with cross campaign drift below 1.5 percent. A separate quality evaluation tests FP16, AWQ 4bit, FP8 weights, and FP8 KV cache on 200 GSM8K questions with five examples per prompt. Sparse attention is evaluated only in simulation. On the calibrated grid, 18 of 36 configurations reach the Pareto frontier. Combined methods reach it more often than individual methods, with 9 of 15 combinations versus 9 of 21 single methods. Quality testing changes the winners. AWQ 4bit reduces per token latency to 0.34 times baseline on L4 but loses 5.9 percent of strict GSM8K accuracy, narrowly missing the 95 percent quality floor within sampling uncertainty. Flexible answer extraction matches FP16 accuracy, suggesting the loss comes from formatting rather than arithmetic. FP8 weights retain 99.4 percent of baseline accuracy at 0.61 to 0.65 times baseline latency across all three GPUs and appear in three of four regime winners. A naive FP8 KV cache maintains normal throughput but answers none of the 200 questions correctly, showing why speed alone is insufficient. Under two prompt designs, n gram speculative decoding measures at 0.90 to 0.98 times baseline and adds no benefit on this stack. The best choice depends on the constraint and GPU: H100 wins for tight latency, while A100 wins for throughput and low cost at 0.106 dollars per million tokens.