Papers with TruthfulQA
Instruction Tuning with Human Curriculum (2024.findings-naacl)
Copied to clipboard
| Challenge: | a recent study shows that human curriculum-inspired strategies can enhance performance of large language models. |
| Approach: | They propose a method for generating instruction-response datasets that emulate human learning . they find that substantial improvements can be achieved through curriculum ordering . |
| Outcome: | The proposed method achieves performance improvements on truthfulQA, MMLU, OpenbookQA, and ARC-hard benchmarks without additional computational costs. |
Alleviating Hallucinations in Large Language Models via Truthfulness-driven Rank-adaptive LoRA (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to improve truthfulness are training-free without modifying the LLM itself. |
| Approach: | They propose a rank-adaptive LoRA method to improve LLM truthfulness that allocates ranks according to truthfulness correlations of LLM modules. |
| Outcome: | The proposed method outperforms state-of-the-art methods on the LLM family and makes the performance of 7B LLMs exceed GPT-4. |
Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation (2024.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to addressing factual inaccuracies require high-quality human factuality annotations to mitigate these hallucinations. |
| Approach: | They propose to leverage the self-evaluation capability of an LLM to provide training signals that steer the model towards factuality. |
| Outcome: | The proposed approach significantly improves factual accuracy over LLMs across three key knowledge-intensive tasks on TruthfulQA and BioGEN. |
When the Model Said ‘No Comment’, We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing work uses SFT and MoE to align Large Language Models, but these work face challenges in multi-objective settings. |
| Approach: | They propose a framework that uses prompt-injected fine-tuning to extract axis-specific task features . it deploys a MoCaE module that calibrates expert routing using fractal and natural geometry . |
| Outcome: | The proposed framework achieves significant gains on Alpaca, BeaverTails, TruthfulQA and TruthfulQ with +171.5% win rate and +110.1% truthfulness-informativeness. |
When Hindsight is Not 20/20: Testing Limits on Reflective Thinking in Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Recent studies suggest that self-reflective prompting can significantly enhance the reasoning capabilities of Large Language Models (LLMs). |
| Approach: | They propose guidelines for when to implement self-reflection in Large Language Models. |
| Outcome: | The proposed approach improves the reasoning capabilities of Large Language Models under a more stringent evaluation setting, and reduces tendency toward majority voting. |
PoLLMgraph: Unraveling Hallucinations in Large Language Models via State Transition Dynamics (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing studies have recognized hallucination as a notable concern in large autoregressive language models (LLMs). |
| Approach: | They propose a polygraph for large language models that detects "hallucination" they demonstrate that hallucination can be detected by tractable probabilistic models . |
| Outcome: | The proposed model outperforms state-of-the-art methods on open-source LLMs by 20% on TruthfulQA benchmarks. |
Sycophancy Hides Linearly in the Attention Heads (2026.eacl-long)
Copied to clipboard
Rifo Ahmad Genadi, Munachiso Samuel Nwadike, Nurdaulet Mukhituly, Tatsuya Hiraoka, Hilal AlQuabeh, Kentaro Inui
| Challenge: | Using TruthfulQA as the base dataset, we find that probes trained on TruthfulQ transfer effectively to other factual QA benchmarks. |
| Approach: | They train linear probes across the residual stream, multilayer perceptron, and attention layers to analyze where sycophancy signals emerge. |
| Outcome: | The proposed model can be used to steer truthfulness and toxicity behaviors. |
Selective Self-to-Supervised Fine-Tuning for Generalization in Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can be fine-tuned on task-specific data to improve performance on target tasks but can be overfitted resulting in a loss of generalization. |
| Approach: | They propose a method that uses the correct model responses from a training set to fine-tune the model using the correct response and the gold response for the remaining samples. |
| Outcome: | The proposed approach reduces model specialization during the fine-tuning stage while improving generalization. |
INVITE: a Testbed of Automatically Generated Invalid Questions to Evaluate Large Language Models for Hallucinations (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in Large language models have enabled them to hold free form conversations over multiple turns, but they exhibit a tendency to make unfounded and incorrect statements, commonly labeled as hallucinations. |
| Approach: | They propose a framework to test large language models for hallucinations using automatically generated INValId questions. |
| Outcome: | The proposed framework is based on a testbed of automatically generated INValId questions to evaluate large language models for hallucinations. |
VeritasQA: A Truthfulness Benchmark Aimed at Multilingual Transferability (2025.coling-main)
Copied to clipboard
Javier Aula-Blasco, Júlia Falcão, Susana Sotelo, Silvia Paniagua, Aitor Gonzalez-Agirre, Marta Villegas
| Challenge: | Large Language Models (LLMs) struggle with falsehoods and model hallucination . many efforts struggle to surpass 50% accuracy, with only targeted techniques reaching around 65% . |
| Approach: | They propose a truthfulness benchmark that focuses on imitative falsehoods . they use a set of 353 questions and answers inspired by common misconceptions based on the language . |
| Outcome: | The benchmark is available in Spanish, Catalan, Galician and English . it measures the truthfulness of multilingual LLMs using 353 questions and answers . |
Alleviating Hallucinations of Large Language Models through Induced Hallucinations (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing studies have shown that large language models generate inaccurate or fabricated information, a phenomenon known as hallucinations. |
| Approach: | They propose a simple strategy to induce-then-contrast decode LLMs to enhance their factuality . they first induce hallucinations from the original model and penalize them . |
| Outcome: | The proposed strategy improves factuality of large language models across task formats, model sizes, and model families. |
Does ChatGPT Know That It Does Not Know? Evaluating the Black-Box Calibration of ChatGPT (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent performance of ChatGPT in downstream tasks is questionable, but does it know that it does not know? |
| Approach: | They propose to use three types of proxy confidence to evaluate ChatGPT's black-box calibration ability. |
| Outcome: | The proposed model exhibits a positive correlation with accuracy in TruthfulQA and a negative correlation in the ModAr dataset. |
XQuant: Achieving Ultra-Low Bit KV Cache Quantization with Cross-Layer Compression (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks. however, their extensive memory requirements present significant challenges for deployment in resource-constrained environments. |
| Approach: | They propose a training-free framework that achieves ultra-low equivalent bit-width KV cache quantization. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on TruthfulQA and LongBench. |
MisinfoBench: A Multi-Dimensional Benchmark for Evaluating LLMs’ Resilience to Misinformation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing benchmarks assess factual accuracy in isolated queries but fail to evaluate LLMs’ resilience to misinformation in interactive settings. |
| Approach: | MisinfoBench is a benchmark designed to assess LLMs’ ability to discern, resist, and reject misinformation. |
| Outcome: | MisinfoBench assesses large language models’ ability to discern, resist, and reject misinformation in interactive settings. |
RLHS: Mitigating Misalignment in RLHF with Hindsight Simulation (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement Learning from Hindsight Simulation (RLHF) can cause severe misalignment in generative AI, but it is not a universal method for fine-tuning large language models. |
| Approach: | They propose a method that uses evaluator feedback to decouple alignment signal from potentially compromised predictions. |
| Outcome: | The proposed method significantly outperforms RLHF in comparisons with baselines and human evaluations. |
Enhancing Language Model Factuality via Activation-Based Confidence Calibration and Guided Decoding (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to calibrate language models are limited in inference-time efficiency or fail to provide informative signals. |
| Approach: | They propose an activation-based calibration method, ActCab, which trains a linear layer on top of the LM’s last-layer activations. |
| Outcome: | The proposed method improves on five popular QA benchmarks and reduces the average expected calibration error (ECE) score by up to 39%. |
SkillAggregation: Reference-free LLM-Dependent Aggregation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods in NLP assign equal weight to all LLM judgments or are designed for specific tasks such as hallucination detection. |
| Approach: | They propose a method that learns to combine LLM judgments without additional data or ground truth to exploit the judge estimates during inference. |
| Outcome: | The proposed method outperforms Crowdlayer on all tasks and yields the best performance over all approaches on the majority of tasks. |
FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for inference-time steering fail to be effective, utility-preserving and training-efficient due to rigid, one-size-fits-all designs and limited adaptability. |
| Approach: | They propose a steering framework that decomposes inference-time steering into two stages . they propose 'conditional steering' mechanism that preserves model utility by avoiding unnecessary steering . a 'mixture-of-Steering-Experts' mechanism captures multimodal nature of desired steering behaviors . |
| Outcome: | The proposed framework outperforms the state-of-the-art methods on safety and truthfulness benchmarks. |
VarBench: Robust Language Model Benchmarking Through Dynamic Variable Perturbation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent benchmarks release only training and validation sets, keeping the test set labels closed-source. |
| Approach: | They propose to extract variables from each test case and define a value range for each variable. |
| Outcome: | The proposed method improves the accuracy of the evaluations on four datasets covering mathematical generation and multiple-choice tasks. |
DeCoVec: Building Decoding Space based Task Vector for Large Language Models via In-Context Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to steering large language models require fine-tuning or manipulation of internal states, limiting their flexibility and scalability. |
| Approach: | They propose a framework that constructs task vectors directly in the decoding space by leveraging in-context learning. |
| Outcome: | The proposed framework outperforms standard few-shot baselines on TruthfulQA, Math-500, and AQUA-RAT with gains up to +5.50 accuracy. |
Pruning Weights but Not Truth: Safeguarding Truthfulness While Pruning LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Neural network pruning disrupts LLMs’ internal activation features crucial for lie detection . layer-wise pruning sparsity inadvertently removes crucial weights, failing to improve lie detection performance despite its reliance on the most crucial LLM layer. |
| Approach: | They propose a pruning approach that places greater emphasis on layers with more activation outliers and stronger discriminative features simultaneously. |
| Outcome: | The proposed approach improves the hallucination detection for pruned LLMs (achieving 88% accuracy at 50% sparsity) and enhances their performance on TruthfulQA. |
SDC-LoRA: Singular-Subspace Drift Controlled LoRA to Mitigate Knowledge Forgetting (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to adapt LLMs to new tasks focus on limiting knowledge forgetting . et al., 2023b) suggest a solution to this problem by limiting update energy in the principal singular subspace of W0 . |
| Approach: | They propose a low-rank Adaptation (LoRA) that steers early updates away from principal directions and mitigates forgetting by constraining update energy in the principal singular subspace of W0. |
| Outcome: | The proposed model mitigates forgetting on MMLU, TruthfulQA, and HellaSwag while keeping minor-subspace updates unchanged. |
SaGE: Evaluating Moral Consistency in Large Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing studies on Large Language Models (LLMs) have focused on accuracy but lack universally agreed-upon answers for moral scenarios. |
| Approach: | They propose a measure called Semantic Graph Entropy to measure a model's moral consistency grounded in "Rules of Thumb" they construct a moral Consistency Corpus (MCC) with 50K moral questions and the RoTs they followed to investigate LLM consistency on two popular datasets. |
| Outcome: | The proposed measure measures moral consistency on two popular datasets . |
Model Unlearning via Sparse Autoencoder Subspace Guided Projections (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing unlearning strategies lack interpretability or fail to provide robust defense against adversarial prompts. |
| Approach: | They propose a framework that leverages SAE features to drive targeted updates in the model’s parameter space. |
| Outcome: | The proposed framework reduces harmful knowledge accuracy by 3.22% compared to baselines and improves adversarial robustness under jailbreak prompts. |
Too Helpful, Too Harmless, Too Honest or Just Right? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods optimize for individual alignment dimensions in isolation, leading to trade-offs and inconsistent behavior. |
| Approach: | They propose a modular alignment framework that integrates a Mixture of Calibrated Experts (MoCaE) within the Transformer architecture. |
| Outcome: | The proposed framework outperforms baselines on three alignment benchmarks, achieving 32.5% win rate, 33.9% safety score, and 28.4% truthfulness. |
Sounding vs. Being an Expert: Disentangling Authority, Register and Cultural Impact in Sycophantic LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models exhibit sycophancy, a tendency to align with user assertions even when they conflict with factual correctness. |
| Approach: | They propose an adversarial evaluation framework that isolates two drivers of credibility: explicit authority (credentials) and implicit authority (linguistic register). |
| Outcome: | The proposed framework disentangles two drivers of credibility: explicit authority (credentials) and implicit authority (linguistic register). |
Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization (2026.findings-acl)
Copied to clipboard
Mingkuan Zhao, Wentao Hu, Tianchen Huang, Yuheng Min, Suquan Chen, Yide Gao, Yanbo Zhai, Shuangyong Song, Xuelong Li
| Challenge: | Hallucinations in Large Language Models persist in critical domains where generated content diverges from contextual facts or logical constraints. |
| Approach: | They propose to generate hallucinations as orthogonal noise relative to the semantic manifold of the residual stream. |
| Outcome: | The proposed method achieves superior contextual faithfulness compared to state-of-the-art methods. |
Reducing Hallucinations in LLMs via Factuality-Aware Preference Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Preference alignment methods can reinforce hallucinations when preference judgments reward fluency and confidence over factual correctness. |
| Approach: | They propose a method that corrects misordered preference pairs and adds a factuality-aware margin to emphasize pairs with clear correctness differences. |
| Outcome: | The proposed method improves factuality and reduces hallucination rates across seven open-weight LLMs. |
GraphSynth: Resolving the Diversity-Reliability Trade-off with Probabilistic Factor Graphs (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models are a scaleable solution for the generation of synthetic data . however, the utility of such data is capped by a critical tension between diversity and factual reliability. |
| Approach: | They propose a framework which leverages a probabilistic factor graph modeling the universe of attributes. |
| Outcome: | The proposed framework outperforms state-of-the-art models with a high structural integrity and a boost in performance on downstream tasks. |
Awakening Dormant Experts:Counterfactual Routing to Mitigate MoE Hallucinations (2026.acl-long)
Copied to clipboard
Wentao Hu, Yanbo Zhai, Xiaohui Hu, Mingkuan Zhao, Shanhong yu, Xue Liu, Kaidong Yu, Shuangyong Song, Xuelong Li
| Challenge: | Sparse Mixture-of-Experts models are vulnerable to hallucinations, authors say . static Top-k routing leaves "specialist experts" under-prioritized for specific tokens . |
| Approach: | They propose a training-free inference framework to awaken dormant experts . they propose 'counterfactual routing' to shift computational resources from syntax-dominant to knowledge-intensive layers . |
| Outcome: | Experiments show that CoR improves factual accuracy by 3.1% without increasing the inference budget. |