Find the smallest machine a task actually needs.

In many real deployments, a large language model is not expected to remain an all-rounder: it repeatedly performs one specific task, such as question answering, summarization, translation, sentiment classification, or natural-language inference. Yet contemporary model compression largely remains model-centric—applying broadly similar pruning or quantization policies regardless of what the deployed model will actually be asked to do. This can preserve computational capacity valuable for general-purpose behavior but unnecessary for the target workload. We investigate a different question: how much of an off-the-shelf LLM is actually required to preserve a single task? Starting directly from the same pretrained open-weight model, without task-specific fine-tuning, we hypothesize that different tasks depend on distinct computational structures and require different numerical precision across those structures. Consequently, the optimal compressed model should be viewed not as a property of the backbone alone, but as a joint property of the model, target task, and deployment budget.

We introduce TASKSCULPT, a framework for Task-Conditioned Computational Surgery that discovers and exploits this structure across multiple architectural resolutions. Given only a task calibration set, TASKSCULPT progressively profiles the model from transformer layers to attention/MLP sub-blocks, attention heads and projections, MLP channel groups, and quantization groups. Rather than assigning importance using a single magnitude heuristic, TASKSCULPT separates inexpensive screening signals from causal ground truth, estimates task-conditioned sensitivity through causal knockout, gradient/Fisher information, activation-aware saliency, and direct compression perturbations, and constructs a task-specific structural rate–distortion function D_t(u,a) for preservation, structured removal, and BF16/FP8/INT8/INT4/INT3/INT2 precision. Selective pairwise interventions probe conditional redundancy: structures that appear removable individually but become critical when compressed together. A closed-loop, hardware-constrained optimizer jointly determines what computation to remove and what precision the surviving computation requires, subject to a strict task-retention constraint. The central causal control is wrong-task map transfer, where a model for one task is compressed using another task’s topology under identical memory, sparsity, calibration, and average-bit budgets.

We plan to evaluate the same off-the-shelf 7–8B-class model across SST-2, MNLI, SQuAD 2.0, XSum, and machine translation, against uniform low-bit quantization, activation- and second-order-aware quantization, generic structural pruning, and recent task-aware pruning and mixed-precision methods. The primary commercial operating point targets ≥99% retention of the original task performance while testing whether TASKSCULPT can reach approximately 6× lower raw weight memory, ~30% hardware-realizable structural removal, ~3.8 average effective bits, and ~2× inference acceleration; a more aggressive frontier targets ≥98% retention. These are explicit engineering targets, not achieved results. The decisive empirical question is whether task-matched compression consistently dominates both task-agnostic policies and wrong-task compression at identical deployment budgets. If confirmed, the implication is a different specialist-LLM deployment paradigm: extract the minimum task-preserving computational subnetwork and precision topology from the stronger general-purpose backbone itself.