Large Language Models (LLMs) are strong at multi-step reasoning, but running them for every generated token is expensive. Small language models (SLMs) are much cheaper, yet less reliable at the hardest reasoning steps. Existing routing methods mostly make a single query-level decision, assuming that response difficulty is uniform once decoding begins. For chain-of-thought reasoning, however, difficulty varies during generation, making token-level routing a more appropriate granularity. This setting introduces three challenges: token-level uncertainty is noisy, local routing decisions must satisfy global efficiency constraints, and frequent switching can disrupt reasoning coherence. We present TokenRouter, a black-box runtime controller for uncertainty-guided token-level routing between small and large language models. TokenRouter combines a robust entropy detector, a self-calibrating control policy, and a dynamic intervention commitment mechanism for stable switching over difficult spans. Across three reasoning benchmarks, TokenRouter maintains over 99% of pure-LLM accuracy while reducing large-model activation to 56.2%–76.0% of generated tokens. Code is available at https://github.com/sznnzs/TokenRouter.

Activation steering controls language models by adding vectors or features to hidden states at inference time, but the upstream source of these steering signals is often treated as a secondary detail. We study this source choice as activation source selection: the combination of source context and activation readout policy used to collect the hidden states from which a steering signal is built. Holding the downstream intervention fixed, we show across three instruction-tuned models and four steering task families that changing only the source activations can more than double steering success. We further find that effective steering is not explained simply by whether the desired behavior appears in the source text. Instead, strong signals come from execution-boundary states, where the model is about to produce or continue the target behavior. This phase distinction explains why answer-based sources sometimes work: their useful component aligns with execution-boundary directions rather than target appearance alone. Building on this view, we introduce tail subtraction, which removes shared prompt and continuation semantics from boundary states and yields cleaner, more stable steering signals. Overall, our results suggest that steering depends on representations of what the model is about to do, not merely on what has already appeared.

Text embeddings in RAG indexes, vector stores, and agent memories are rarely leaked as isolated vectors: they often appear with document IDs, chunk offsets, and preserved order. This paper shows that such ordered provenance is not passive metadata but an active privacy channel. We introduce MIRAGE, a public-shadow, zero-victim-pair attack that exploits ordered document-chunk leakage by turning facts recovered from earlier chunks into encoder-verified anchors for later reconstruction. MIRAGE builds public-shadow single-chunk candidates, extracts typed anchors, verifies them with counterfactual encoder-support checks, and uses verified memory to regenerate document-consistent chunks. Across five dense encoders and four decoders on ordered Enron and synthetic CanaryDocs, under a strict 512-query victim-time budget per chunk, MIRAGE outperforms strong fixed-bank reranking, improving macro privacy-unit recall by 5.65-6.10 points across encoders and 5.2-6.9 points across decoders; gains persist (+3.7-4.4 points) on two real, non-templated corpora. Its support check raises accepted-anchor precision from 43.6% to 69.4%, and 34.8% of newly recovered privacy units appear only in regenerated candidates absent from every fixed bank. These results reveal document grouping, chunk order, and offsets as sensitive attack surfaces that RAG and vector-store deployments must protect alongside embeddings themselves.

KV cache compression is critical for long-context LLM inference, yet existing training-free methods fall into a dilemma: attention-based eviction clusters retained tokens around semantically “hot” regions, creating context holes that degrade long-range reasoning, while diversity-based selection retains directionally broad but semantically irrelevant tokens. We identify the root cause: at prefill time, future decoding queries are unknown, making query-conditioned signals unreliable as a sole selection criterion. The retained KV set must instead provide broad directional coverage of the context—so that arbitrary future queries can find supporting keys—while filtering out tokens that expand the coverage span without contributing to generation. We propose AnchorKV, which realizes this principle through block-wise partitioning with uniform per-block budgets and anchor-driven cosine dissimilarity scoring for linear-time diversity estimation. A compression-rate adaptive hybrid score dynamically reweights diversity and importance. Experiments on LongBench across three model families show AnchorKV consistently outperforms existing training-free methods, with the largest margins at low retention ratios. A query-subsampled variant reduces prefill latency by up to 2.7× with minimal accuracy loss. We provide theoretical guarantees on approximation error and variance reduction.

Large Reasoning Models (LRMs) show remarkable capabilities in complex reasoning tasks through Chain-of-Thought (CoT) mechanisms. However, their reasoning processes frequently suffer from overthinking, generating inefficient patterns and substantial overhead without improving quality. Existing pruning approaches typically rely on computationally expensive external teacher models or aggregated sequence-level metrics, failing to capture the fine-grained, intrinsic cognitive state of individual reasoning steps. In this paper, we propose ThinkLeap, a novel confidence-guided optimization framework that identifies and retains only critical cognitive leaps within reasoning chains. Our approach is grounded in the observation that a model’s internal confidence dynamics serve as a fine-grained structural signal: rising confidence often reflects evidence accumulation, while declining confidence signals doubt or redundancy. By leveraging these dynamics, ThinkLeap autonomously replaces inefficient steps with ellipsis markers, preserving the logical flow while signaling omitted deliberation. Extensive experiments across three challenging reasoning benchmarks demonstrate that ThinkLeap establishes a new paradigm for efficient reasoning, achieving an average 19.4% reduction in reasoning length while maintaining comparable accuracy.

Large Language Models (LLMs) face significant memory and latency overheads during long-context inference due to the growing KV cache, especially in Knowledge Base Question Answering (KBQA) settings that require support for multiple downstream queries. Query-aware eviction methods do not generalize across queries, while existing query-agnostic approaches rely on a single proxy query, leading to fragile eviction decisions under high eviction ratios. We propose ContrastKV, a robust query-agnostic KV cache eviction algorithm for multi-query generalization. ContrastKV introduces a contrastive signal fusion mechanism that jointly exploits complementary semantic and non-semantic signals. By contrasting semantic consistency with structural robustness, the method constructs a more reliable eviction criterion that alleviates the blind spots of single-query proxies. The framework integrates efficient signal generation, parallel importance scoring, and multi-level fusion across heads and layers. Experiments show that ContrastKV outperforms state-of-the-art methods, retaining up to 92% accuracy with only 20% of the KV cache budget, while reducing decoding latency by approximately 50% and significantly lowering GPU memory usage.

Large Reasoning Models (LRMs) demonstrate strong performance on complex tasks through chain-of-thought (CoT) reasoning. However, they suffer from high inference latency due to lengthy reasoning chains. In this paper, we propose SpecCoT, a collaborative framework that combines large and small models for effective yet efficient reasoning. Unlike traditional speculative decoding, which operates at the token level, SpecCoT adopts a step-level verification strategy: the large model first establishes the reasoning direction, and for each intermediate step, the small model generates multiple candidate drafts in parallel. The large model then verifies these drafts, either selecting the most suitable one or rejecting them all and generating its own. SpecCoT approach balances reasoning quality with inference efficiency through fine-grained model cooperation. Experiments across diverse tasks show SpecCoT reduces inference latency by 1.7-4.1× while maintaining comparable accuracy to standard large model inference.

Recent advances in large language models (LLMs) have demonstrated remarkable potential in the field of natural language processing. Unfortunately, LLMs face significant security and ethical risks. Although techniques such as safety alignment are developed for defense, prior researches reveal the possibility of bypassing such defenses through well-designed jailbreak attacks. In this paper, we propose QueryAttack, a novel framework to examine the generalizability of safety alignment. By treating LLMs as knowledge databases, we translate malicious queries in natural language into structured non-natural query language to bypass the safety alignment mechanisms of LLMs. We conduct extensive experiments on mainstream LLMs, and the results show that QueryAttack not only can achieve high attack success rates (ASRs), but also can jailbreak various defense methods. Furthermore, we tailor a defense method against QueryAttack, which can reduce ASR by up to 64% on GPT-4-1106. Our code is available at https://anonymous.4open.science/r/QueryAttack-334B.

Large language models (LLMs) excel at complex reasoning tasks but often suffer from overconfidence and computational inefficiency due to fixed computation budgets and miscalibrated confidence estimates. We present a novel framework for computationally efficient, trustworthy reasoning under uncertainty, introducing two complementary techniques: Diversity-Aware Self-Signal Dilution (DASD) and Convergent Adaptive Weighted Sampling (CAWS). DASD operates in an unsupervised manner to dilute overconfident, semantically redundant reasoning paths, thereby producing better-calibrated internal confidence estimates. CAWS dynamically allocates computational resources at inference time by aggregating these signals and terminating computation once answer dominance and stability are achieved. Comprehensive experiments across three reasoning datasets demonstrate that our approach maintains accuracy levels while achieving over 70% reduction in inference cost, surpassing competitive baselines. Our framework provides a scalable, unsupervised solution for reliable and efficient LLM reasoning.

Aligned Large Language Models (LLMs) showcase remarkable versatility, capable of handling diverse real-world tasks. Meanwhile, aligned LLMs are also expected to exhibit speciality, excelling in specific applications. However, fine-tuning with extra data, a common practice to gain speciality, often leads to catastrophic forgetting (CF) of previously acquired versatility, hindering the model’s performance across diverse tasks. In response to this challenge, we propose CoFiTune, a coarse to fine framework in an attempt to strike the balance between speciality and versatility. At the coarse-grained level, an empirical tree-search algorithm is utilized to pinpoint and update specific modules that are crucial for speciality, while keeping other parameters frozen; at the fine-grained level, a soft-masking mechanism regulates the update to the LLMs, mitigating the CF issue without harming speciality. In an overall evaluation of both speciality and versatility, CoFiTune consistently outperforms baseline methods across diverse tasks and model scales. Compared to the full-parameter SFT, CoFiTune leads to about 14% versatility improvement and marginal speciality loss on a 13B model. Lastly, based on further analysis, we provide a speculative insight into the information forwarding process in LLMs, which helps explain the effectiveness of the proposed method. The code is available at https://github.com/rattlesnakey/CoFiTune.