Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations “flip”, such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model’s instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces.

Quantization enables efficient deployment of large language models (LLMs) in resource-constrained environments by significantly reducing memory and computation costs. While quantized LLMs often maintain performance on perplexity and zero-shot tasks, their impact on truthfulness—whether generating truthful or deceptive responses—remains largely unexplored. In this work, we introduce TruthfulnessEval, a comprehensive evaluation framework for assessing the truthfulness of quantized LLMs across three dimensions: (1) Truthfulness on Logical Reasoning; (2) Truthfulness on Common Sense; and (3) Truthfulness on Imitative Falsehoods. Using this framework, we examine mainstream quantization techniques (ranging from 4-bit to extreme 2-bit) across several open-source LLMs. Surprisingly, we find that while quantized models retain internally truthful representations, they are more susceptible to producing false outputs under misleading prompts. To probe this vulnerability, we test 15 rephrased variants of “honest”, “neutral” and “deceptive” prompts and observe that “deceptive” prompts can override truth-consistent behavior, whereas “honest” and “neutral” prompts maintain stable outputs. Further, we reveal that quantized models “know” the truth internally yet still produce false outputs when guided by “deceptive” prompts via layer-wise probing and PCA visualizations. Our findings provide insights into future designs of quantization-aware alignment and truthfulness interventions.

The efficacy of Large Language Models (LLMs) in long-context tasks is often hampered by the substantial memory footprint and computational demands of the Key-Value (KV) cache. Current compression strategies, including token eviction and learned projections, frequently lead to biased representations—either by overemphasizing recent/high-attention tokens or by repeatedly degrading information from earlier context—and may require costly model retraining. We present FAEDKV (Frequency-Adaptive Infinite-Window for KV cache), a novel, training-free KV cache compression framework that ensures unbiased information retention. FAEDKV operates by transforming the KV cache into the frequency domain using a proposed Infinite-Window Fourier Transform (IWDFT). This approach allows for the equalized contribution of all tokens to the compressed representation, effectively preserving both early and recent contextual information. A preliminary frequency ablation study identifies critical spectral components for layer-wise, targeted compression. Experiments on LongBench benchmark demonstrate FAEDKV’s superiority over existing methods by up to 22%. In addition, our method shows superior, position-agnostic retrieval accuracy on the Needle-In-A-Haystack task compared to compression based approaches.

Neural network pruning has emerged as a promising approach for deploying LLMs in low-resource scenarios while preserving downstream task performance. However, for the first time, we reveal that such pruning disrupts LLMs’ internal activation features crucial for lie detection, where probing classifiers (typically small logistic regression models) trained on these features assess the truthfulness of LLM-generated statements. This discovery raises a crucial open question: how can we prune LLMs without sacrificing these critical lie detection capabilities? Our investigation further reveals that naively adjusting layer-wise pruning sparsity based on importance inadvertently removes crucial weights, failing to improve lie detection performance despite its reliance on the most crucial LLM layer. To address this issue, we propose Truthful Pruning aligned by Layer-wise Outliers (TPLO), which places greater emphasis on layers with more activation outliers and stronger discriminative features simultaneously. This preserves LLMs’ original performance while retaining critical features of inner states needed for robust lie detection. Moreover, we introduce a prompting rule to enrich the TruthfulQA benchmark for better calibrating LLM pruning. Empirical results show that our approach improves the hallucination detection for pruned LLMs (achieving 88% accuracy at 50% sparsity) and enhances their performance on TruthfulQA.

Large language models (LLMs) have achieved promising results in tabular data generation. However, inherent historical biases in tabular datasets often cause LLMs to exacerbate fairness issues, particularly when multiple advantaged and protected features are involved. In this work, we introduce a universal debiasing framework that minimizes group-level dependencies by simultaneously reducing the mutual information between advantaged and protected attributes. By leveraging the autoregressive structure and analytic sampling distributions of LLM-based tabular data generators, our approach efficiently computes mutual information, reducing the need for cumbersome numerical estimations. Building on this foundation, we propose two complementary methods: a direct preference optimization (DPO)-based strategy, namely UDF-DPO, that integrates seamlessly with existing models, and a targeted debiasing technique, namely UDF-MIX, that achieves debiasing without tuning the parameters of LLMs. Extensive experiments demonstrate that our framework effectively balances fairness and utility, offering a scalable and practical solution for debiasing in high-stakes applications.

Aspect Sentiment Triplet Extraction (ASTE) is one of the compound tasks of fine-grained aspect-based sentiment analysis (ABSA), aiming at extracting the triplets of aspect terms, corresponding opinion terms and the associated sentiment orientation. Recent efforts in exploiting span-level semantic interaction shown superior performance on ASTE task. However, most of the existing span-based approaches suffer from enumerating all possible spans, since it can introduce too much noise in sentiment triplet extraction. To ease this burden, we propose a dual-channel span generation method to coherently constrain the search space of span candidates. Specifically, we leverage the syntactic relations among aspect/opinion terms and the associated part-of-speech characteristics in those terms to generate span candidates, which reduces span enumeration by nearly half. Besides, feature representations are learned from syntactic and part-of-speech correlation among terms, which renders span representation fruitful linguistic information. Extensive experiments on two versions of public datasets demonstrate both the effectiveness of our design and the superiority on ASTE/ATE/OTE tasks.

Understanding the semantic meaning of content on the web through the lens of entities and concepts has many practical advantages. However, when building large-scale entity extraction systems, practitioners are facing unique challenges involving finding the best ways to leverage the scale and variety of data available on internet platforms. We present learnings from our efforts in building an entity extraction system for multiple document types at large scale using multi-modal Transformers. We empirically demonstrate the effectiveness of multi-lingual, multi-task and cross-document type learning. We also discuss the label collection schemes that help to minimize the amount of noise in the collected data.

In this paper, we propose a novel model RevGAN that automatically generates controllable and personalized user reviews based on the arbitrarily given sentimental and stylistic information. RevGAN utilizes the combination of three novel components, including self-attentive recursive autoencoders, conditional discriminators, and personalized decoders. We test its performance on the several real-world datasets, where our model significantly outperforms state-of-the-art generation models in terms of sentence quality, coherence, personalization, and human evaluations. We also empirically show that the generated reviews could not be easily distinguished from the organically produced reviews and that they follow the same statistical linguistics laws.