Mathematical word problems with plausible irrelevant context expose a failure mode in which language models reason from facts that should not support the target solution. We formulate distractor-robust mathematical reasoning as relevance-boundary supervision: a model should separate Signal facts from Distractor facts before solving, and then ground the solution in Signal facts only. We introduce ReBound, a post-training framework that constructs paired clean–distractor examples, trains boundary-first traces with supervised fine-tuning, and applies correctness-guided reinforcement learning to discourage Distractor use. Across three open base LLMs and six mathematical reasoning benchmarks, ReBound consistently improves distractor robustness over CoT prompting and strong training-based baselines; on Qwen3-0.6B, the distractor benchmark average increases from 34.92 to 48.29. ReBound also preserves competitive clean and perturbation performance. Process diagnostics, ablations, and output-length analysis show that these gains are tied to better fact use and boundary behavior, rather than longer generations or answer-only optimization.The code is available at https://anonymous.4open.science/r/Rebound-5FB8

Column Type Annotation (CTA) enriches tables with semantic column types that are often missing or incomplete in real-world data, aiding multiple downstream table tasks. Yet CTA remains challenging because sparse or noisy column values often require table-level context for disambiguation. Prior work either flattens tables into encoder inputs, losing structure and semantics, or follows a description-then-encoding pipeline that separates encoding and adds generation cost. We propose D2R, a decoder-based framework that obtains column representations directly from the hidden state that initiates column description generation. Specifically, D2R first encodes the table context once as a shared prefix KV cache, providing table-level context for all target columns. For each column, we append its sampled values and a description trigger token, then extract the trigger’s hidden state as the column embedding, preserving semantics without producing text. A decoder-attached classifier, trained via contrastive learning aligned with the CTA objective, predicts column types directly from this embedding. Across four CTA benchmarks, D2R achieves the best Micro-F1, outperforming TURL, a strong table PLM, by up to 12.57 points, while retaining efficient inference. The code is available at https://github.com/IvesCheung/cta.

Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token dominance. This overhead is driven by a multi-regime interplay between high-resolution feature extraction, quadratic attention scaling, and memory bandwidth constraints. We present a systematic taxonomy of efficiency techniques structured around the inference lifecycle, consisting of encoding, prefilling, and decoding. Unlike prior reviews focused on isolated optimizations, we analyze the end-to-end pipeline to reveal how upstream decisions dictate downstream bottlenecks, covering compute-bound visual encoding, the intensive prefilling of massive contexts, and the ”visual memory wall” in bandwidth-bound decoding. By decoupling the efficiency landscape into the axes of shaping information density, managing long-context attention, and overcoming memory limits, this work provides a structured analysis of how isolated optimizations compose to navigate the trade-off between visual fidelity and system efficiency. The survey concludes by outlining four future frontiers supported by pilot empirical insights, including hybrid compression based on functional unit sensitivity, modality-aware decoding with relaxed verification, progressive state management for streaming continuity, and stage-disaggregated serving through hardware-algorithm co-design. The submitted software contains a snapshot of our literature repository, which is designed to be maintained as a living resource for the community.

Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key–value (KV) caches. Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs. A common solution is to compress caches under a fixed allocated budget at different granularities: token-level uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads. Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies. We propose HybridKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval. Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7B show that HybridKV reduces KV cache memory by up to 7.9× and achieves 1.52× faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.

Domain adaptation is widely adopted in text retrieval scenarios where large labeled data is unavailable. To improve model adaptability, existing methods try to expand more source datasets. However, we found from experiments that indiscriminately using a large amount of source data from various text tasks does not guarantee improved adaptability, but may negatively impact model performance. To tackle this issue, we propose Trait, a framework that can effectively improve model adaptability by selecting beneficial data without evaluating all source data. Specifically, we first divide multiple source datasets into data chunks of the same size as the minimum selection unit to form the whole selection space. Then we devise an iterative process that includes Bayesian optimization-based selection and transfer-aware chunk evaluation to incrementally select beneficial chunks. To reduce unnecessary evaluation costs, we also design backtracking and pruning actions to adjust the selection subspace. Extensive experimental results show that Trait not only achieves average state-of-the-art for few-shot on nine target datasets by evaluating only 4% of BERRI source data, but also is very competitive for zero-shot compared with LLM-based rankers.

In this paper, we propose SkipBERT to accelerate BERT inference by skipping the computation of shallow layers. To achieve this, our approach encodes small text chunks into independent representations, which are then materialized to approximate the shallow representation of BERT. Since the use of such approximation is inexpensive compared with transformer calculations, we leverage it to replace the shallow layers of BERT to skip their runtime overhead. With off-the-shelf early exit mechanisms, we also skip redundant computation from the highest few layers to further improve inference efficiency. Results on GLUE show that our approach can reduce latency by 65% without sacrificing performance. By using only two-layer transformer calculations, we can still maintain 95% accuracy of BERT.