Papers with inference overhead
What the Router Sees Matters: Funnel Pooling for Fast, Content Driven Expert Routing (2026.eacl-srw)
Copied to clipboard
| Challenge: | Modern large language model (LLM) systems often route inputs to specialized experts to improve accuracy, efficiency, and robustness. |
| Approach: | They propose a lightweight trainable in-model readout that constructs the routing vector directly from token-level hidden states. |
| Outcome: | The proposed model reduces language and source-dataset driven clustering and results in more topic-aligned domains. |
From Schema to State: Zero-Shot Scheme-Only Dialogue State Tracking via Diverse Synthetic Dialogue and Step-by-Step Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing research classifies zero-shot, scheme-only DST into two main types: the cross-domain scenario and the zero-schemaonly setting. |
| Approach: | They propose a zero-shot, scheme-only approach that generates synthetic dialogues that balance diversity with schema alignment and distills knowledge from a large language model into a smaller model. |
| Outcome: | The proposed approach achieves state-of-the-art performance under zero-shot, scheme-only situation and generalizes effectively to few-shot scenarios. |
AdaptThink: Reasoning Models Can Learn When to Think (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in large reasoning models have demonstrated remarkable capabilities in tackling complex tasks. |
| Approach: | They propose an algorithm to teach reasoning models to choose the optimal thinking mode based on problem difficulty. |
| Outcome: | The proposed algorithm reduces the average response length and improves accuracy on three math datasets. |
BaseCal: Unsupervised Confidence Calibration via Base Model Signals (2026.acl-long)
Copied to clipboard
Hexiang Tan, Wanli Yang, Junwei Zhang, Xin Chen, Rui Tang, Du Su, Jingang Wang, Yuanzhuo Wang, Fei Sun, Xueqi Cheng
| Challenge: | Post-trained LLMs typically compromise reliability with severe overconfidence, resulting in inaccurate responses. |
| Approach: | They propose a solution that feeds PoLLMs into the base LLM to get confidence. |
| Outcome: | The proposed solution reduces expected calibration error (ECE) by 42.90% compared to the best unsupervised baselines. |
Efficient Nearest Neighbor based Uncertainty Estimation for Natural Language Processing Tasks (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods for uncertainty estimation are inadequate for safety-critical applications. |
| Approach: | They propose a method that uses the distances from neighbors and the ratio of labels in neighbors to estimate uncertainty. |
| Outcome: | The proposed method outperforms baseline and density-based methods in calibration and uncertainty metrics. |
SuLoRA: Subspace Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning large language models (LLMs) introduce parameter interference, leading to a gap in generalization performance for specific tasks compared to full fine-uning. |
| Approach: | They propose a parameter-separated low-rank adapter to account for task differences by decomposing LoRA’s parameter matrix into multiple independent subspaces and assigning them differentially to distinct tasks. |
| Outcome: | The proposed method outperforms LoRA in trainable parameter efficiency and overall model performance on various NLP tasks. |
CoT-Valve: Length-Compressible Chain-of-Thought Tuning (2025.acl-long)
Copied to clipboard
| Challenge: | Wei et al., 2022) have developed a powerful method for enhancing the reasoning capabilities of large language models. |
| Approach: | They propose to use a tuning and inference strategy to control the length of reasoning chains by a parameter space direction to control their length. |
| Outcome: | The proposed method reduces reasoning chains on GSM8K from 741 to 225 tokens with a minor performance drop (95.07% to 94.92%) and on AIME from 6827 to 4629 tokens, with only one additional incorrect answer. |
Safeguarding Language Models via Self-Destruct Trapdoor (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing mechanisms to restrict behavior of language models (LMs) are vulnerable to misuse and misalignment. |
| Approach: | They propose a mechanism to restrict specific behaviors in language models by exploiting hardware properties. |
| Outcome: | The proposed mechanism can be applied to trigger overflows for specific behaviors or target hardware malfunctions. |
BERM: Training the Balanced and Extractable Representation for Matching to Improve Generalization Ability of Dense Retrieval (2023.acl-long)
Copied to clipboard
| Challenge: | Dense retrieval has shown promise in the first-stage retrieval process when trained on in-domain labeled datasets. |
| Approach: | They propose a method to capture matching signal to improve generalization of dense retrieval by capturing matching signal between two texts. |
| Outcome: | The proposed method can be combined with different training methods to improve generalization ability without additional inference overhead and target domain data. |
Self-Distillation for Model Stacking Unlocks Cross-Lingual NLU in 200+ Languages (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) excel on English NLU tasks, yet struggle to extend their NLU capabilities to underrepresented languages. |
| Approach: | They integrate machine translation models (MT) directly into LLM backbones via sample-efficient self-distillation. |
| Outcome: | The proposed model outperforms translation-test models on 127 low-resource languages. |
Decoding in Latent Spaces for Efficient Inference in LLM-based Recommendation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Light Latent-space Decoding (L2D) is an efficient and efficient latent- space decoding method. |
| Approach: | They propose to bypass language-space decoding by matching candidate items with LLM's internal thought representations in the latent space. |
| Outcome: | The proposed method is 10x faster than language-space decoding while maintaining or enhancing performance. |
IIET: Efficient Numerical Transformer via Implicit Iterative Euler Method (2025.emnlp-main)
Copied to clipboard
Xinyu Liu, Bei Li, Jiahao Liu, Junhao Ruan, Kechen Jiao, Hongyin Tang, Jingang Wang, Tong Xiao, JingBo Zhu
| Challenge: | High-order numerical methods enhance performance in tasks like NLP but introduce a performance-efficiency trade-off due to increased computational overhead. |
| Approach: | They propose an iterative implicit Euler Transformer which simplifies high-order numerical methods by iterating implicit Eule. |
| Outcome: | The proposed method improves accuracy and reduces inference overhead by 55% while retaining 99.4% of the original task accuracy. |
IntactKV: Improving Large Language Model Quantization by Keeping Pivot Tokens Intact (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing quantization methods are compromising performance of large language models (LLMs) despite their high computational intensity, LLMs are still demanding intensive computation. |
| Approach: | They propose to generate the KV cache of pivot tokens losslessly from the full-precision model. |
| Outcome: | The proposed method generates the KV cache of pivot tokens losslessly from the full-precision model with no extra inference overhead. |
Efficient Nearest Neighbor Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Non-parametric neural language models (NLMs) learn text distributions by memorizing training data points. |
| Approach: | They propose to use an external datastore to learn from a non-parametric language model. |
| Outcome: | The proposed methods achieve up to a 6x speed-up in inference speed while retaining comparable performance. |
Think Twice, Generate Once: Safeguarding by Progressive Self-Reflection (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models generate coherent and contextually relevant text, but their deployment raises significant concerns about the potential for harmful or inappropriate content. |
| Approach: | They propose a novel inference-time technique that empowers LLMs to self-monitor and correct their outputs dynamically. |
| Outcome: | The proposed method reduces the attack success rate from 77.47% to 5.86%, to Llama-3.1-8B base from 89.70% to 5.56%, and to Qwen2.5-7B-Instruct from 44.44% to 3.84%, without additional training. |
MoQAE: Mixed-Precision Quantization for Long-Context LLM Inference via Mixture of Quantization-Aware Experts (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to optimize large language models for long-context inference are inefficient and consume memory. |
| Approach: | They propose a mixed-precision quantization method via mixture of experts that inputs tokens into router chunk by chunk to reduce inference overhead. |
| Outcome: | The proposed method outperforms state-of-the-art KV cache quantization methods on multiple benchmark datasets. |
Learning Contextual Retrieval for Robust Conversational Search (2025.emnlp-main)
Copied to clipboard
| Challenge: | Effective conversational search requires a deep understanding of user intent across multiple dialogue turns. |
| Approach: | They propose a novel LLM-based retriever that directly incorporates conversational context into the retrieval process. |
| Outcome: | The proposed method outperforms existing methods while incurring no additional inference overhead. |
Nudging: Inference-time Alignment of LLMs via Guided Decoding (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) require alignment to effectively and safely follow user instructions. |
| Approach: | They propose a simple, training-free algorithm that aligns any base model at inference time using a small aligned model. |
| Outcome: | The proposed algorithm outperforms large aligned models on open-instruction tasks without training. |
Exploring the Limitations of Mamba in COPY and CoT Reasoning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Inference overhead of Transformers increases linearly with the sequence length, posing challenges for modeling long sequences. |
| Approach: | They analyze Mamba's expressive ability to perform COPY operations and Chain of Thought reasoning tasks using a defined sequence length. |
| Outcome: | The proposed model can perform COPY operations and Chain of Thought reasoning tasks with a constant size while reducing computational costs. |
AMIA: Automatic Masking and Joint Intention Analysis Makes LVLMs Robust Jailbreak Defenders (2025.findings-emnlp)
Copied to clipboard
| Challenge: | AMIA is a lightweight, inference-only defense for Large Vision–Language Models . it automatically masks text-irrelevant image patches and conducts joint Intention Analysis . |
| Approach: | AMIA is a lightweight, inference-only defense for large vision–language models . it automatically masks a small set of text-irrelevant image patches to disrupt adversarial perturbations . |
| Outcome: | AMIA improves defense success rates across diverse LVLMs and jailbreak benchmarks . it preserves general utility with only 2% accuracy drop, incurs only modest inference overhead . |
You Only Need One Single Token to Refine Safety Alignment (2026.findings-acl)
Copied to clipboard
| Challenge: | Excessive safety can lead to over-refusal, where models reject harmful-looking yet benign queries, severely limiting utility. |
| Approach: | They propose a lightweight training-based approach that reshapes the distributions of harmful and benign samples within the model’s decision space by using a single-token prefix. |
| Outcome: | The proposed approach can distinguish between harmful and benign samples while keeping the model frozen. |
Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated exceptional proficiency in language-related tasks, but their deployment poses significant memory and storage requirements. |
| Approach: | They propose a method that optimizes rounding values and weight clipping within 200 steps. |
| Outcome: | The proposed method achieves exceptional results across 2 to 4 bits while maintaining low tuning costs and avoiding additional inference overhead. |
O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning (2026.findings-acl)
Copied to clipboard
Haotian Luo, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Xiaochun Cao, Dacheng Tao, Naiqiang Tan, Li Shen
| Challenge: | Recent long-thought reasoning models adopt extended reasoning processes similar to how humans ponder over complex problems. |
| Approach: | They propose a model that uses RL-style fine-tuning to reduce inference overhead while maintaining accuracy. |
| Outcome: | The proposed model reduces inference overhead while maintaining accuracy. |
VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies evaluate efficiency robustness of vision-language models under unrealistic assumptions, requiring access to model architecture and parameters. |
| Approach: | They propose a novel approach to evaluate VLM efficiency robustness in a realistic black-box setting. |
| Outcome: | The proposed approach generates adversarial images with imperceptible perturbations, increasing the computational cost by up to 128.47%. |
TARE: Lightweight Token-Aware Representation Editing for Fine-tuning Transformer-like Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing PEFT methods can be costly and underfit token-level contexts. |
| Approach: | They propose a PEFT method that performs fine-grained, token-specific edits with a small additional inference overhead and minimal tuning. |
| Outcome: | The proposed method outperforms state-of-the-art methods in 8 tasks and GLUE with a minimal tuning overhead and inference overhead. |
AdaTooler-V: Adaptive Tool-Use for Images and Videos (2026.findings-acl)
Copied to clipboard
Chaoyang Wang, Kaituo Feng, Dongyang Chen, Zhongyu Wang, Zhixun Li, Sicheng Gao, Meng Meng, Xu Zhou, Manyuan Zhang, Yuzhang Shang, Xiangyu Yue
| Challenge: | Existing models exhibit blind tool-use reasoning patterns, which significantly increases inference overhead and degrades model performance. |
| Approach: | They propose an MLLM that performs adaptive tool-use by determining whether a visual problem truly requires tools. |
| Outcome: | The proposed model outperforms existing methods in visual reasoning tasks. |
Adaptive Zooming via Relevance-Informed Positional Resource Allocation for Training-free LLM Context Extension (2026.findings-acl)
Copied to clipboard
Hongbo Zhao, Huibin Wang, Bin Tang, Xianming Hu, Yihong Huang, Yijun Shen, Nuoyi Chen, Ping Li, Kai Zhang
| Challenge: | Existing training-free methods for extrapolating beyond training context lengths are semantics-agnostic . Existing methods that focus on relative token distances can indiscriminately blur semantically relevant and irrelevant tokens . |
| Approach: | They propose an adaptive positional zooming method that uses semantic relevance to extrapolate beyond training context lengths. |
| Outcome: | Experiments show that RiPRA outperforms existing training-free extrapolation methods . relevant tokens get higher positional resolution, while irrelevant tokens are compressed . |
RACC: Regret-Aware Confidence Calibration for Consistent Masked Discrete Diffusion Decoding (2026.findings-acl)
Copied to clipboard
| Challenge: | Masked Discrete Diffusion Models (MDMs) enable parallel generation via iterative refinement, but their current decoding paradigms are static and myopic. |
| Approach: | They propose a Regret-Aware Confidence Calibration framework that aligns decoding decisions with the model’s latent self-correction capabilities. |
| Outcome: | The proposed framework aligns decoding decisions with model’s latent self-correction capabilities. |
GER-LLM: Efficient and Effective Geospatial Entity Resolution with Large Language Model (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for integrating spatial data from diverse sources are limited by their reliance on large amounts of training data and their inability to incorporate commonsense knowledge. |
| Approach: | They propose a framework that integrates large language models into the GER pipeline. |
| Outcome: | The proposed framework improves on real-world geospatial datasets and shows that it is more efficient than state-of-the-art methods. |
Grouped Adaptive Weight Sharing (GAWS): An Inference-Efficient Adaptation Method for Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Low-Rank Adaptation (LoRA) is a new approach to fine-tuning large language models . adapters are lightweight, task specific modules that can be used for adapters in latency-sensitive settings. |
| Approach: | They propose a low-rank adapter with a weight sharing mechanism that reduces latency by 40% . they analyze LoRA adapters on GPUs and identify segmented function calls as the primary source of latency. |
| Outcome: | The proposed adapter reduces latency to about 40% of the gap between the unmerged LoRA and the base model while maintaining parameter efficiency and comparable accuracy. |