Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters. Moreover, it integrates seamlessly into downstream alignment pipelines, highlighting its potential for efficient alignment.

The widespread deployment of Large Language Models (LLMs) has spurred significant progress in the detection of LLM-generated text.However, existing detection methods often rely on statistical features that are insufficient for reliable detection; for example, even though LLM-generated and humanwritten texts exhibit different probability distributions in surrogate models, they can produce nearly identical entropy values, thereby conflating the two types of text.In this paper, we propose that modulating the decoding temperature and monitoring how the probability distributions respond can better probe the intrinsic discrepancies between two types of text.Building upon this insight, we introduce a new feature termed Temperature Sensitivity (TS) and demonstrate that LLM-generated text tends to exhibit higher TS than humanwritten text.Finally, we propose NTS, a novel and simple zero-shot detector built upon normalized temperature sensitivity.Extensive experiments across three datasets, multiple domains, and various source models demonstrate the superior effectiveness and robustness of our proposed approach.Code avaliable at

Large Language Models (LLMs) suffer from hallucinations, severely undermining their reliability.While white-box hallucination detection methods that leverage hidden states prevail, they fail to identify and focus on factcritical information when analyzing token sequences.To address this, we propose LAFaCT, a Localize-then-Analyze detection framework.It first localizes fact-critical tokens using Factual Criticality, a novel metric derived from feature attribution.A subsequent stage then performs a focused sequential analysis on their hidden states.Extensive experiments on eight benchmarks and multiple model families confirm LAFaCT as the new state-of-the-art, with in-depth analyses validating the effectiveness of its core token-localization strategy.

In-context learning (ICL), which promotes inference with several demonstrations, has become a widespread paradigm to stimulate LLM capabilities for downstream tasks.Due to context length constraints, it cannot be further improved in spite of more training data, and general features directly from LLMs in ICL are not adaptive to the specific downstream task.In this paper, we propose a feature-adaptive and datascalable in-context learning framework (FADS-ICL), which can leverage task-adaptive features to promote inference on the downstream task, with the supervision of beyond-context samples.Specifically, it first extracts general features of beyond-context samples via the LLM with ICL input form one by one, and introduces a task-specific modulator to perform feature refinement and prediction after fitting a specific downstream task.We conduct extensive experiments on FADS-ICL under varying data settings (4∼128 shots) and LLM scale (0.8∼70B) settings.Experimental results show that FADS-ICL consistently outperforms previous state-of-the-art methods by a significant margin under all settings, verifying the effectiveness and superiority of FADS-ICL.For example, under the 1.5B and 32 shots setting, FADS-ICL can achieve +14.3 average accuracy from feature adaptation over vanilla ICL on 10 datasets, with +6.2 average accuracy over the previous state-of-the-art method, and the performance can further improve with increasing training data.

Chinese spelling check (CSC) is a fundamental NLP task that detects and corrects spelling errors in Chinese texts.As most of these spelling errors are caused by phonetic similarity, effectively modeling the pronunciation of Chinese characters is a key factor for CSC.In this paper, we consider introducing an auxiliary task of Chinese pronunciation prediction (CPP) to improve CSC, and, for the first time, systematically discuss the adaptivity and granularity of this auxiliary task.We propose SCOPE which builds on top of a shared encoder two parallel decoders, one for the primary CSC task and the other for a fine-grained auxiliary CPP task, with a novel adaptive weighting scheme to balance the two tasks.In addition, we design a delicate iterative correction strategy for further improvements during inference.Empirical evaluation shows that SCOPE achieves new state-of-theart on three CSC benchmarks, demonstrating the effectiveness and superiority of the auxiliary CPP task.Comprehensive ablation studies further verify the positive effects of adaptivity and granularity of the task.Code and data used in this paper are publicly available at https: //github.com/jiahaozhenbang/SCOPE.