Baoliang Tian

Ning Yang, Hengyu Zhong, Wentao Wang et al.

The extension of context windows in Large Language Models is typically facilitated by scaling positional encodings followed by lightweight Continual Pre-Training (CPT). While effective for processing long sequences, this paradigm often disrupts original model capabilities, leading to performance degradation on standard short-text benchmarks. We propose LinearARD, a self-distillation method that restores Rotary Position Embeddings (RoPE)-scaled students through attention-structure consistency with a frozen native-RoPE teacher. Rather than matching opaque hidden states, LinearARD aligns the row-wise distributions of dense $Q/Q$, $K/K$, and $V/V$ self-relation matrices to directly supervise attention dynamics. To overcome the quadratic memory bottleneck of $n \times n$ relation maps, we introduce a linear-memory kernel. This kernel leverages per-token log-sum-exp statistics and fuses logit recomputation into the backward pass to compute exact Kullback-Leibler divergence and gradients. On LLaMA2-7B extended from 4K to 32K, LinearARD recovers 98.3\% of the short-text performance of state-of-the-art baselines while surpassing them on long-context benchmarks. Notably, our method achieves these results using only \textbf{4.25M} training tokens compared to the \textbf{256M} tokens required by LongReD and CPT. Our code is available at https://github.com/gracefulning/LinearARD.

Ning Yang, Chengzhi Wang, Yibo Liu et al.

Large Language Models (LLMs) in long-context scenarios are severely constrained by the linear growth of Key-Value (KV) cache memory. Existing KV compression methods rely either on static thresholds and attention-only heuristics or on coarse memory budget allocation. Under tight memory budgets, these methods overlook two key factors: prompt-dependent variation in compression risk and functional heterogeneity across attention heads, which destabilize token selection and lead to tail failures. To address these challenges, we propose CompilerKV, a risk-adaptive and head-aware compression framework that compiles offline experience into reusable decision tables for prefill-only deployment. CompilerKV integrates two key synergistic components: (i) a Head Heterogeneity Table, learned via offline contextual bandits, which assigns head-specific reliability weights to govern functional differences across attention heads explicitly; and (ii) a Risk-Adaptive Threshold Gating mechanism that jointly models attention entropy and local perplexity, transforming prompt-level risk into deployable retention thresholds. Experiments on LongBench show CompilerKV dominates SOTA methods under a 512-token budget, recovering 97.7\% of FullKV performance while achieving up to +5.2 points gain over the strongest competitor.

Dongxu Zhang, Ning Yang, Jihua Zhu et al.

Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Large Language Models (LLMs), yet the reliability of these reasoning chains remains a critical challenge. A widely held "cascading failure" hypothesis suggests that errors are most detrimental when they occur early in the reasoning process. This paper challenges that assumption through systematic error-injection experiments, revealing a counter-intuitive phenomenon we term "Late-Stage Fragility": errors introduced in the later stages of a CoT chain are significantly more likely to corrupt the final answer than identical errors made at the beginning. To address this specific vulnerability, we introduce the Adaptive Self-Correction Chain-of-Thought (ASCoT) method. ASCoT employs a modular pipeline in which an Adaptive Verification Manager (AVM) operates first, followed by the Multi-Perspective Self-Correction Engine (MSCE). The AVM leverages a Positional Impact Score function I(k) that assigns different weights based on the position within the reasoning chains, addressing the Late-Stage Fragility issue by identifying and prioritizing high-risk, late-stage steps. Once these critical steps are identified, the MSCE applies robust, dual-path correction specifically to the failure parts. Extensive experiments on benchmarks such as GSM8K and MATH demonstrate that ASCoT achieves outstanding accuracy, outperforming strong baselines, including standard CoT. Our work underscores the importance of diagnosing specific failure modes in LLM reasoning and advocates for a shift from uniform verification strategies to adaptive, vulnerability-aware correction mechanisms.

Ning Yang, Hai Lin, Yibo Liu et al.

Ensuring that large language models (LLMs) generate outputs aligned with human preferences is important for safe and effective AI interactions. While Direct Preference Optimization (DPO) employs an implicit reward function to optimize the policy model, however, it and its related variants overlook the differential importance of individual tokens and are sensitive to judgment noise in preference datasets during generation. Although recent methods attempt to assess the important weight of tokens via probability prediction or simplistic weighting schemes, these evaluation methods are prone to biases and still cannot fully address these issues. To solve this problem, we propose the Token-Importance Guided Direct Preference Optimization (TI-DPO), which introduces two key innovations: the gradient-based token-importance weights that dynamically prioritize critical tokens, and a triple loss that explicitly guides model outputs to approach human-preferred responses and stay away from non-preferred responses. Experimental results show that TI-DPO achieves higher accuracy and stronger generative diversity, providing more stable and computationally efficient solutions compared with DPO and other RLHF methods.