{ "schema_version": 1, "builder_plan": { "scenario": "A self-hosted model uses more compute or memory than your workload can justify.", "change": "Check for a usable pruned checkpoint and runtime support before attempting model surgery.", "baseline": "The original unpruned model at the same precision, context, and load.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Task quality, actual GPU memory, serving throughput, and preparation cost.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2610.08341", "title": "DIPrune: Task-Aware Token Pruning with Dual Importance for Efficient Multimodal Language Models", "summary": "DIPrune replaces task-agnostic or attention-based training-free visual token pruning in MLLMs with a dual importance score combining intra-layer feature saliency and inter-layer semantic evolution derived from a task-loss distortion bound. It costs additional per-layer scoring and ranking computation at inference and implementation complexity, though it requires no retraining and reduces visual token count and thus compute on a single GPU. It can fail if the gradient-based inter-layer importance estimates are unreliable for a given task distribution, or if aggressive pruning ratios still discard tokens needed for deep reasoning despite the corrective term.", "rationale": "Training-free token pruning directly fits a single-GPU budget and can cut inference cost, but the claimed gains are validated only on LLaVA and Qwen-VL benchmarks, so the reader should validate accuracy retention on their own workload before production adoption.", "technique": "Task-aware visual token pruning for MLLMs", "family": "esparsidade_e_poda", "recommendation": "testar", "published": "2026-10-06", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.08341", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "DIPrune replaces task-agnostic or attention-based training-free visual token pruning in MLLMs with a dual importance score combining intra-layer feature saliency and inter-layer semantic evolution derived from a task-loss distortion bound. It costs additional per-layer scoring and ranking computation at inference and implementation complexity, though it requires no retraining and reduces visual token count and thus compute on a single GPU. It can fail if the gradient-based inter-layer importance estimates are unreliable for a given task distribution, or if aggressive pruning ratios still discard tokens needed for deep reasoning despite the corrective term.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.08341", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "The paper states DIPrune 'consistently achieves state-of-the-art results' on LLaVA and Qwen-VL among training-free pruning methods; no specific factor or accuracy delta is given in the abstract.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.08341", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }