training-free

Nullify 提出一种免训练、非破坏性的激活引导方法用于 LLM 知识遗忘，在推理时通过满足零空间约束的引导向量重定向隐私相关激活，在 TOFU 和 MUSE 基准上匹配或超越现有基线，同时近乎无损地保留模型效用。

The paper proposes KITE, a training-free framework that shifts latent multi-agent KV communication from sender-side state fidelity to receiver-side task sufficiency, transmitting only a task-effective key layer. Experiments show 28-36× communication reduction, up to 3× inference speedup, and accuracy gains of up to 23.3 percentage points across seven benchmarks.

The paper introduces KVFetch, a training-free framework that adds a temporal recall channel to score-based KV cache compressors to fix "sequential forgetting" (verbatim copying breaking under compression), recovering copying performance from 0.8 to 78.4 and boosting the RULER-16K average by +8.4 at iso-budget.

论文提出 CoDR（Confidence-Drift Remasking），一种免训练、与采样器无关的后处理精炼方法，通过检测已提交 token 的置信度漂移并对其进行重遮蔽与重新生成，从而在掩码扩散语言模型的推理与编码任务上以较少的前向传播次数提升准确率。

The paper analyzes the coverage-selectivity trade-off of global unsafety subspaces in training-free text-to-image safeguards and proposes CALM, a prompt-local counterfactual correction method that improves unsafe content suppression while preserving benign utility.

The paper introduces MintFlow, a training-free constrained sampling framework for flow matching models that enforces constraints via minimal, closed-form perturbations of intermediate flow states, preserving the pretrained generative distribution better than state-of-the-art constrained methods.

提出 MC-Sparse，一个免训练的元缓存稀疏注意力框架，通过分析稠密-稀疏注意力差距的三大来源，在视频与3D生成中实现最高2.32倍去噪加速且几乎无质量损失。

HLA-WM提出一种训练免费的混合线性注意力框架，通过几何引导的检索与递归线性状态重组合来缓解GatedDeltaNet在长时视频世界模型中的远距离遗忘问题，在60秒基准上以12倍内存节省同时提升所有一致性与相机控制指标。

This paper provides a mechanistic explanation for why diffusion language model agents fall into retry loops after failed actions, attributing it to a task-blind corruption in masked decoding, and proposes Reflect Reverse, a training-free method that scores actions via the reverse conditional of the task given the state, improving success rates on four embodied benchmarks.

The paper proposes EB-GAD, a training-free framework that reframes node-level graph anomaly detection as finite-horizon control, fitting a graph precision via Empirical Bayes and scoring anomalies with closed-form control energy. It achieves the best or tied-best AUROC on 9 of 11 benchmarks without using labels at any step.

The paper introduces FORESIGHT, a training-free dual-stream architecture using two Siamese LLMs to anticipate future context and dynamically reconfigure computation in streaming vision-language models. With a frozen Qwen3-VL-8B backbone, it achieves 23.0 mean joint F1 on OmniPro Online and improves StreamingBench by 6.7 and OVO-Bench by 15.4 over the base model.

Periscope introduces a training-free inference method that arranges text chunks on a K×K grid and queries a frozen language model over local and strided spans to recover answers beyond the context window, matching or beating full-window reads on LongBench v2 and InfiniteBench at a fraction of the cost. The approach lets a 27B model read 4.5M-token contexts on a single 80GB GPU by replacing long reads with an evidence map.

The paper introduces LoopCD, a training-free contrastive decoding framework for looped transformers that uses intermediate recurrent states as weak/strong prediction pairs, delivering substantial gains (e.g., Ouro-2.6B-Thinking's AIME 2024 pass@1 rising from 61.88% to 73.33%) while enabling halving of recurrent loops and 22.5%-48.2% reductions in forward FLOPs.

ReGain 提出一种训练无关的采样时校正方法，通过按频率带缩放被放大的 CFG 引导，恢复基于合成图像微调的个性化扩散模型的主体保真度，在 SD1.5、SDXL 和 SD3.5 上均有效。

NEEDLE is a training-free method for removing backdoors from LLMs that estimates a backdoor direction and a refusal subspace, then applies sequential weight orthogonalisation to suppress backdoors while preserving refusal behaviour and model capability. It achieves the lowest Attack Success Rate (including 0% on code injection attacks) with minimal KL divergence and safety degradation across multiple model families.

The paper introduces EmoRES-TTS, a training-free method for emotional speech generation that enhances controllability by decomposing emotion vectors into shared and residual components, achieving superior performance over existing methods on benchmarks.

JevSoup introduces a training-free framework for routing and combining LoRA experts without additional training, achieving performance gains over existing methods on multiple tasks.

This paper proposes a training-free pipeline for pronunciation transcription that integrates lexical resources and acoustic models to achieve low error rates and high efficiency, outperforming baselines including commercial multimodal LLMs.

QuantWM 提出一种免训练的 2-bit KV cache 量化框架，通过量化敏感性感知聚类（QSAC）与主子空间注意力补偿（PSAC）来稳定视频世界模型的注意力 logits，缓解时间闪烁与视觉退化，并在多个基准上实现最高 6.20 倍 KV cache 内存压缩。

提出了 ExpActivator，一个免训练框架，通过在冻结的 GUI 骨干网络的潜在空间中匹配当前屏幕与历史状态，仅激活与当前情境相关的经验，而非盲目追加任务相关历史；在四个 GUI 骨干网络上将轨迹内步骤成功率平均提升 28%，同时仅使用约五分之一的历史 token。