Tag
ARIA is a test-time unlearning method for large language models that uses sparse autoencoders to suppress unwanted knowledge during inference without modifying weights, improving the forget-retain trade-off and remaining robust to adversarial attacks.
The paper introduces ReActNet, a training-free framework that synthesizes task-conditioned temporal workflow graphs for multi-agent LLM systems, enhancing coordination and performance over fixed or learned topology baselines.
ERASE is a framework for functional forgetting in machine learning models without modifying weights, using input perturbations to suppress private data influence during inference for privacy compliance.
This paper introduces Private Best-of-N (PrivBoN) and Private Inference-Time Pessimism (PrivITP) methods that add calibrated noise to reward scores in inference-time alignment to achieve differential privacy and mitigate reward hashing, with minimal additional alignment cost.
This paper proposes EmoVec, a lightweight framework for controllable affective generation in large language models via latent vector steering, enabling continuous control over emotional intensity without model weight updates.
CritICL is an inference-time framework that enhances large language model reasoning by leveraging structured failure patterns from smaller models as critique-based guidance, outperforming standard in-context learning with reduced generation and token costs.
GGSS reduces demographic bias in generative vision-language models by steering visual tokens along geodesic arcs with an adaptive gate during inference, preserving visual-language accuracy.
The paper introduces Readout Feedback (RoFB), a test-time intervention that steers latent dynamics in recurrent reasoning models using their own readout probabilities, achieving performance gains on Sudoku and Maze tasks without retraining.
The paper introduces 'recirculation', an inference-time architectural enhancement for foundation models that markedly reduces perplexity and boosts accuracy by improving state tracking, with no additional latency during generation.
This paper proposes inference-time mitigation strategies using Chain of Thought prompting and Direct Preference Optimization to reduce adversarial political bias in large language models, demonstrating significant improvements in political neutrality scores.
This paper introduces a question-level audit framework distinguishing 'realized' from 'reachable' answers on LLM benchmarks, showing that aggregate score gains often come from producing already-reachable answers rather than expanding true capability, and that random layer routing matches structured search under matched budgets.
This paper proposes an inference-time policy shaping framework to steer pretrained reinforcement learning policies toward welfare-based fairness objectives without retraining, inspired by inference-time alignment in LLMs.
Introduces V-Steer, a training-free inference-time method that edits cached value vectors to restore instruction hierarchy in language models, raising primary constraint accuracy from under 18% to 92% on controlled benchmarks with negligible overhead.
Proposes DiFA, a training-free framework that reframes inference-time data prediction refinement as sequential state estimation using Kalman filtering, significantly improving generative fidelity on CIFAR-10 and ImageNet.
Motion4Motion is a training-free motion transfer framework that models motion flow from a video instead of relying on skeleton structures, enabling motion transfer across diverse species (e.g., humans to animals) at inference time.
Proposes LARA, a framework for safe inference-time alignment that uses Lagrangian dualization to derive an augmented reward from separate reward and cost models, improving the helpfulness-harmlessness tradeoff without retraining.
This paper reveals that hallucination in large vision-language models is caused by a dynamic structural misalignment where certain attention heads act as risky mediators, decoupling from visual evidence to lock onto language priors. The authors propose Fox, a training-free causal intervention framework that diagnoses and physically severs these pathological shortcuts, achieving state-of-the-art performance in faithful decoding.
Proposes G-SPIN, a lightweight framework that combines phonetic graph modeling with contextual language understanding for correcting ASR errors, using a GNN to generate phonetically plausible candidate tokens, an MLM for local scoring, and an LLM for final re-ranking, all operating at inference time.
This paper introduces a conversational voice agent system that uses a lightweight on-device 'Talker' model to start responding immediately, then incorporates knowledge from a frontier LLM 'Reasoner' as it becomes available, achieving 7-19x faster time-to-first-response while approaching frontier-level performance on a laptop.
Proposes ARIADNE, a training-free, adapter-agnostic routing framework that selects the optimal PEFT adapter at inference time by measuring input proximity to adapter-specific centroids in embedding space, recovering 97.44% of upper-bound performance on 23 tasks.