Aligning Large Language Models (LLMs) to be faithful to new knowledge in complex, multi-hop reasoning tasks is a critical, yet unsolved, challenge. We find that SFT-based methods, e.g., Reason-KE, while state-of-the-art, suffer from a "faithfulness gap": they optimize for format mimicry rather than sound reasoning. This gap enables the LLM’s powerful parametric priors to override new contextual facts, resulting in critical factual hallucinations (e.g., incorrectly reasoning "Houston" from "NASA" despite an explicit edit). To solve this core LLM alignment problem, we propose **Reason-KE++**, an SFT+RL framework that instills process-level faithfulness. Its core is a Stage-aware Reward mechanism that provides dense supervision for intermediate reasoning steps (e.g., Decomposition, Sub-answer Correctness). Crucially, we identify that naive outcome-only RL is a deceptive trap for LLM alignment: it collapses reasoning integrity (e.g., 19.00% Hop acc) while superficially boosting final accuracy. Our process-aware framework sets **a new SOTA of 95.48%** on MQUAKE-CF-3k (+5.28%), demonstrating that for complex tasks, aligning the reasoning process is essential for building trustworthy LLMs.

Recently, long-thought reasoning LLMs, such as OpenAI’s O1, adopt extended reasoning processes similar to how humans ponder over complex problems. This reasoning paradigm significantly enhances the model’s problem-solving abilities and achieves promising results. However, long-thought reasoning process leads to a substantial increase in inference time. A pressing challenge is reducing the inference overhead of long-thought LLMs while ensuring accuracy. In this paper, we identify that long-thought reasoning models struggle to effectively allocate token budgets based on problem difficulty and reasoning redundancies. To address this, we propose Length-Harmonizing Fine-Tuning (O1-Pruner), aiming at minimizing reasoning overhead while maintaining accuracy. This effective fine-tuning method first estimates the LLM’s baseline performance through pre-sampling and then uses RL-style fine-tuning to encourage the model to generate shorter reasoning processes under accuracy constraints. This allows the model to achieve efficient reasoning with lower redundancy while maintaining accuracy. Experiments on various mathematical reasoning benchmarks show that O1-Pruner not only significantly reduces inference overhead but also achieves higher accuracy, providing a novel and promising solution to this challenge.

Zeroth-Order optimization presents a promising memory-efficient paradigm for fine-tuning Large Language Models by relying solely on forward passes. However, its practical adoption is severely constrained by slow wall-clock convergence and high estimation variance. In this work, we dissect the runtime characteristics of ZO algorithms and identify a critical system bottleneck where the generation of perturbations and parameter updates accounts for over 40% of the training latency. We argue that the standard uniform exploration strategy is fundamentally flawed as it fails to account for the heterogeneous sensitivity of layers in deep networks, resulting in computationally wasteful blind searches. To address this structural mismatch, we propose **AdaLeZO**, an **Ada**ptive **L**ayer-wis**e** **ZO** optimization framework. By formulating the layer selection process as a non-stationary Multi-Armed Bandit problem, AdaLeZO dynamically allocates the limited perturbation budget to the most sensitive parameters.We further introduce an Inverse Probability Weighting mechanism based on sampling with replacement, which guarantees unbiased gradient estimation while effectively acting as a temporal denoiser to reduce variance. Extensive experiments on LLaMA and OPT models ranging from 6.7B to 30B parameters demonstrate that AdaLeZO achieves 1.7× to 3.0× wall-clock acceleration compared to state-of-the-art methods. Crucially, AdaLeZO functions as a universal plug-and-play module that seamlessly enhances the efficiency of existing ZO optimizers without incurring additional memory overhead.

Fine-tuning-as-a-service, while commercially successful for Large Language Model (LLM) providers, exposes models to harmful fine-tuning attacks. As a widely explored defense paradigm against such attacks, unlearning attempts to remove malicious knowledge from LLMs, thereby essentially preventing them from being used to perform malicious tasks. However, we highlight a critical flaw: the inherent general adaptability of LLMs allows them to easily bypass selective unlearning by rapidly relearning or repurposing their general capabilities for harmful tasks. To address this fundamental limitation, we propose a paradigm shift: instead of selective removal, we advocate for inducing model collapse, effectively forcing the model to ”unlearn everything”, specifically in response to updates characteristic of malicious adaptation. This collapse directly neutralizes the very general capabilities that attackers exploit, tackling the core issue unaddressed by selective unlearning. We introduce the Collapse Trap (CTRAP) as a practical mechanism to implement this concept conditionally. Embedded during alignment, CTRAP pre-configures the model’s reaction to subsequent fine-tuning dynamics. If updates during fine-tuning constitute a persistent attempt to reverse safety alignment, the pre-configured trap triggers a progressive degradation of the model’s core language modeling abilities, ultimately rendering it inert and useless for the attacker. Crucially, this collapse mechanism remains dormant during benign fine-tuning, ensuring the model’s utility and general capabilities are preserved.

Resource constraints often limit the parameter capacity of Large Language Models (LLMs), thereby hindering their performance. Although existing approaches leverage parameter sharing to reuse a fixed set of parameters within constrained budgets, they typically require each layer to fulfill multiple roles over a fixed number of iterations. This design compromises both efficiency and adaptability. In this work, we propose the **Zero Token Transformer (ZTT)**, which employs a head-tail decoupled parameter cycling strategy. Specifically, we decouple the first (head) and last (tail) layers from the parameter cycling process, enabling iterative refinement solely within the intermediate layers. Furthermore, we introduce a Zero-Token Mechanism, wherein a virtual token with a trainable key and a zero-valued vector functions as a standard token. The resulting attention scores not only reflect the computational significance of each layer but also facilitate dynamic early exiting, thereby preserving overall model accuracy. Our approach achieves superior performance under strict parameter constraints, substantially reduces computational overhead via early exits, and can be seamlessly integrated into the fine-tuning of existing pre-trained models, improving both efficiency and adaptability.

Efficiently managing the KV cache in Large Language Models (LLMs) is a critical challenge for long-context processing tasks such as retrieval-augmented generation (RAG), long text summarization, and multi-document analysis. Extending the context length substantially increases the KV cache size, leading to excessive memory consumption. Existing KV cache compression methods enforce a fixed pattern, neglecting task-specific characteristics, which hampers the effective retention of essential information while discarding less important tokens. In this paper, we introduce a novel Task-Aware KV cache mechanism that dynamically adjusts the KV cache size across different layers based on the characteristics of the tasks. Our approach builds on the significant observation of distinct activation patterns across layers in various tasks, which highlights the need for adaptive strategies tailored to each task’s unique demands. Based on this insight, we propose DynamicKV, a method that dynamically optimizes token retention by adjusting the number of tokens retained at each layer, adapting to the specific task. DynamicKV establishes global and per-layer maximum KV cache budgets, temporarily retaining the maximum budget for the current layer, and periodically updating the KV cache sizes of all preceding layers during inference. Our method demonstrates exceptional performance on the LongBench dataset, retaining only 1.7% of the KV cache while preserving 90%, 87%, 78%, and 83% of the original accuracy for LlaMA-3-8B-Instruct, Mistral-7B-Instruct-v0.2, Qwen2-7B-Instruct, and InternLM-2.5-7B-Chat-1M, respectively. When the retained KV cache size is increased to 6.9%, the performance becomes nearly indistinguishable from that without any KV cache compression. Notably, even under extreme compression (0.9%), DynamicKV surpasses state-of-the-art (SOTA) methods by 11% in the Needle-in-a-Haystack test using Mistral-7B-Instruct-v0.2. The code is available at repository https://github.com/DreamMr/DynamicK.

Large language models (LLMs) encode vast amounts of world knowledge but remain static once trained, making timely integration of emerging facts prohibitively expensive via full retraining. Knowledge-editing techniques have thus emerged to inject or overwrite specific facts into LLMs, yet they either over-rely on superficial cues or incur complex, iterative pipelines that collapse under noisy, multi-hop conditions. We introduce **Reason-KE**, an end-to-end reasoning-chain-based editing framework that steers a pretrained LLM through four structured stages—fact acknowledgment, relevance determination, selective application, and final reasoning—to filter distractors in a single pass. Trained on MQuAKE-CF with up to four irrelevant facts, Reason-KE elevates Qwen2.5-7B’s multi-hop QA accuracy to 90.2% (↑17.6 pp) while suffering merely 6.3% drop under heavy distraction and <1% when answers are leaked. Our quantitative analysis confirms Reason-KE’s resilience and efficiency, establishing a new state of the art for reliable LLM knowledge updates. The code will be released.