论文检索

General Machine Learning · Representation Learning

Shauli Ravfogel, Gilad Yehudai, Tal Linzen, Joan Bruna, Alberto Bietti

Recent probing studies reveal that large language models exhibit linear subspaces that separate true from false statements, yet the mechanism behind their emergence is unclear. We introduce a transparent, one-layer transformer toy model that reproduces such truth subspaces end-to-end and exposes one concrete route by which they can arise. We study one simple setting in which truth encoding can emerge: a data distribution where factual statements co-occur with other factual statements (and vice-versa), encouraging the model to learn this distinction in order to lower the LM loss on future tokens. We corroborate this pattern with experiments in pretrained language models. Finally, in the toy setting we observe a two-phase learning dynamic: networks first memorize individual factual associations in a few steps, then---over a longer horizon---learn to linearly separate true from false, which in turn lowers language-modeling loss. Together, these results provide both a mechanistic demonstration and an empirical motivation for how and why linear truth representations can emerge in language models.

shengming yuan, Xinyu Lyu, Shuailong Wang, Beitao Chen, Jingkuan Song, Lianli Gao

Multimodal large language models (MLLMs) face an inherent trade-off between faithfulness and creativity, as different tasks require varying degrees of associative reasoning. However, existing methods lack the flexibility to modulate this reasoning strength, limiting MLLMs' adaptability across factual and creative scenarios. To bridge this gap, we propose equipping MLLMs with mechanisms that enable flexible control over associative reasoning. We begin by investigating the internal mechanisms underlying associative behavior in MLLMs and find that: (1) middle layers play a pivotal role in shaping model’s associative tendencies, (2) modifying representations in these layers effectively regulates associative reasoning strength, and (3) hallucinations can be exploited to derive steering vectors that guide this modulation. Building on these findings, we introduce Flexible Association Control (FlexAC), a lightweight and training-free framework for modulating associative behavior in MLLMs. FlexAC first induces hallucination-guided intermediate representations to encode associative directions. Then, it selects high-association instances to construct effective associative steering vectors, whose strengths are adaptively calibrated to balance creative guidance with output stability. Finally, recognizing the multi-dimensional nature of associative reasoning, FlexAC incorporates task-specific associative vectors derived from a forward pass on a few target-domain samples, enabling models to follow diverse associative directions and better adapt to creative tasks. Notably, our method achieves up to a 5.8× improvement in creativity on Creation-MMBench and a 29\% reduction in hallucination rate on CHAIR, surpassing existing baselines and demonstrating its effectiveness in enabling flexible control over associative reasoning in MLLMs. Our code is available at https://github.com/ylhz/FlexAC.

Deep Learning · Algorithms

Chenyu Zhu, Yefeng Liu, Hao Zhang, Aowen Wang, Yangxue, Guanhua Chen, Longyue Wang, Weihua Luo, Kaifu Zhang

Despite their outstanding performance in numerous applications, large language models (LLMs) remain prone to hallucinations, generating content inconsistent with their pretraining corpora. Currently, almost all contrastive decoding approaches alleviate hallucinations by introducing a model susceptible to hallucinations and appropriately widening the contrastive logits gap between hallucinatory tokens and target tokens. However, although existing contrastive decoding methods mitigate hallucinations, they lack enough confidence in the factual accuracy of the generated content. In this work, we propose Multi-Model Contrastive Decoding (MCD), which integrates a pretrained language model with an evil model and a truthful model for contrastive decoding. Intuitively, a token is assigned a high probability only when deemed potentially hallucinatory by the evil model while being considered factual by the truthful model. This decoding strategy significantly enhances the model’s confidence in its generated responses and reduces potential hallucinations. Furthermore, we introduce a dynamic hallucination detection mechanism that facilitates token-by-token identification of hallucinations during generation and a tree-based revision mechanism to diminish hallucinations further. Extensive experimental evaluations demonstrate that our MCD strategy effectively reduces hallucinations in LLMs and outperforms state-of-the-art methods across various benchmarks.

Computer Vision · Vision Models & Multimodal

Wei Chen, Xin Yan, Bin Wen, Fan Yang, Tingting Gao, Di ZHANG, Long Chen

Although multimodal large language models (MLLMs) exhibit remarkable reasoning capabilities on complex multimodal understanding tasks, they still suffer from the notorious 'hallucination' issue: generating outputs misaligned with obvious visual or factual evidence. Currently, training-based solutions, like direct preference optimization (DPO), leverage paired preference data to suppress hallucinations. However, they risk sacrificing general reasoning capabilities due to the likelihood displacement. Meanwhile, training-free solutions, like contrastive decoding, achieve this goal by subtracting the estimated hallucination pattern from a distorted input. Yet, these handcrafted perturbations (e.g., add noise to images) may poorly capture authentic hallucination patterns. To avoid these weaknesses of existing methods, and realize ``robust'' hallucination mitigation (\ie, maintaining general reasoning performance), we propose a novel framework: Decoupling Contrastive Decoding (DCD). Specifically, DCD decouples the learning of positive and negative samples in preference datasets, and trains separate positive and negative image projections within the MLLM. The negative projection implicitly models real hallucination patterns, which enables vision-aware negative images in the contrastive decoding inference stage. Our DCD alleviates likelihood displacement by avoiding pairwise optimization and generalizes robustly without handcrafted degradation. Extensive ablations across hallucination benchmarks and general reasoning tasks demonstrate the effectiveness of DCD, \ie, it matches DPO’s hallucination suppression while preserving general capabilities and outperforms the handcrafted contrastive decoding methods.

Computer Vision · Everything Else

Ankan Deria, Adinath Dukre, feilong tang, Sara Atito, Sudipta Roy, Muhammad Awais, Muhammad Haris Khan, Imran Razzak

Despite significant advances in inference-time search for vision–language models (VLMs), existing approaches remain both computationally expensive and prone to unpenalized, low-confidence generations which often lead to persistent hallucinations. We introduce \textbf{Value-guided Inference with Margin-based Reward (ViMaR)}, a two-stage inference framework that improves both efficiency and output fidelity by combining a temporal-difference value model with a margin-aware reward adjustment. In the first stage, we perform a single pass to identify the highest-value caption among diverse candidates. In the second stage, we selectively refine only those segments that were overlooked or exhibit weak visual grounding, thereby eliminating frequently rewarded evaluations. A calibrated margin-based penalty discourages low-confidence continuations while preserving descriptive richness. Extensive experiments across multiple VLM architectures demonstrate that ViMaR generates captions that are significantly more reliable, factually accurate, detailed, and explanatory, while achieving over 4$\times$ speedup compared to existing value-guided methods. Specifically, we show that ViMaR trained solely on LLaVA Mistral-7B \textit{generalizes effectively to guide decoding in stronger unseen models}. To further validate this, we adapt ViMaR to steer generation in both LLaVA-OneVision-Qwen2-7B and Qwen2.5-VL-3B, leading to consistent improvements in caption quality and demonstrating robust cross-model guidance. This cross-model generalization highlights ViMaR's flexibility and modularity, positioning it as a scalable and transferable inference-time decoding strategy. Furthermore, when ViMaR-generated captions are used for self-training, the underlying models achieve substantial gains across a broad suite of visual comprehension benchmarks, underscoring the potential of fast, accurate, and self-improving VLM pipelines. Code: https://github.com/ankan8145/ViMaR

Yiyou Sun, Yu Gai, Lijie Chen, Abhilasha Ravichander, Yejin Choi, Nouha Dziri, Dawn Song

Large language models (LLMs) frequently generate hallucinations—content that deviates from factually inaccurate or deviates from provided context—posing challenges for diagnosis. However, diagnosing the causes of hallucination is challenging due to the complex interplay of underlying causes. This paper introduces a framework to systematically understand the sources of hallucination behavior in large language models. Our key insight is that hallucinations arise when more frequent but non-factual associations outweigh faithful ones. Through theoretical and empirical analyses, we demonstrate that decoder-only transformers effectively function as subsequence embedding models, with the fully-connected layers encoding input-output associations. We propose a tracing algorithm that identifies causal subsequences by analyzing hallucination probabilities across randomized input contexts. Experiments show our method outperforms standard attribution techniques in identifying hallucination causes and is supported by evidence from the model’s training corpus. This work provides a unified perspective on hallucinations and a robust framework for their cause and analysis.

Social Aspects · Trustworthy Machine Learning

Yuchen Wu, Edward Sun, Kaijie Zhu, Jianxun Lian, José Hernández-Orallo, Aylin Caliskan, Jindong Wang

Large language models (LLMs) typically generate identical or similar responses for all users given the same prompt, posing serious safety risks in high-stakes applications where user vulnerabilities differ widely. Existing safety evaluations primarily rely on context-independent metrics—such as factuality, bias, or toxicity—overlooking the fact that the same response may carry divergent risks depending on the user's background or condition. We introduce ``personalized safety'' to fill this gap and present PENGUIN—a benchmark comprising 14,000 scenarios across seven sensitive domains with both context-rich and context-free variants. Evaluating six leading LLMs, we demonstrate that personalized user information significantly improves safety scores by 43.2%, confirming the effectiveness of personalization in safety alignment. However, not all context attributes contribute equally to safety enhancement. To address this, we develop RAISE—a training-free, two-stage agent framework that strategically acquires user-specific background. RAISE improves safety scores by up to 31.6% over six vanilla LLMs, while maintaining a low interaction cost of just 2.7 user queries on average. Our findings highlight the importance of selective information gathering in safety-critical domains and offer a practical solution for personalizing LLM responses without model retraining. This work establishes a foundation for safety research that adapts to individual user contexts rather than assuming a universal harm standard.

Computer Vision · Vision Models & Multimodal

Sibei Yang, Ge Zheng, Jiajin Tang, Jiaye Qian, Hanzhuo Huang, Cheng Shi

Large language models (LLMs) and vision-language models (LVLMs) have driven the paradigm shift towards general-purpose foundation models. However, both of them are prone to hallucinations, which compromise their factual accuracy and reliability. While existing research primarily focuses on isolated textual- or visual-centric errors, a critical yet underexplored phenomenon persists in LVLMs: Even neither of textual- or visual centric errors occur, LVLMs often struggle with a new and subtle hallucination mode that arising from composition of them. In this paper, we define this issue as Simple Compositional Hallucination (SCHall). Through an preliminary analysis, we present two key findings: (1) visual abstraction fails under compositional questioning, and (2) visual inputs induce degradation in language processing, leading to hallucinations. To facilitate future research on this phenomenon, we introduce a custom benchmark, SCBench, and propose a novel VLR-distillation method, which serves as the first baseline to effectively mitigate SCHall. Furthermore, experiment results on publicly available benchmarks, including both hallucination-specific and general-purpose ones, demonstrate the effectiveness of our VLR-distillation method.

General Machine Learning · Causality

Peng Wu, Haoxuan Li, Chunyuan Zheng, Yan Zeng, Jiawei Chen, Yang Liu, Ruocheng Guo, Kun Zhang

Counterfactual inference aims to estimate the counterfactual outcome at the individual level given knowledge of an observed treatment and the factual outcome, with broad applications in fields such as epidemiology, econometrics, and management science. Previous methods rely on a known structural causal model (SCM) or assume the homogeneity of the exogenous variable and strict monotonicity between the outcome and exogenous variable. In this paper, we propose a principled approach for identifying and estimating the counterfactual outcome. We first introduce a simple and intuitive rank preservation assumption to identify the counterfactual outcome without relying on a known structural causal model. Building on this, we propose a novel ideal loss for theoretically unbiased learning of the counterfactual outcome and further develop a kernel-based estimator for its empirical estimation. Our theoretical analysis shows that the rank preservation assumption is not stronger than the homogeneity and strict monotonicity assumptions, and shows that the proposed ideal loss is convex, and the proposed estimator is unbiased. Extensive semi-synthetic and real-world experiments are conducted to demonstrate the effectiveness of the proposed method.

Deep Learning · Generative Models and Autoencoders

Hanyu Zhu, Lance Fiondella, Jiawei Yuan, Kai Zeng, Long Jiao

Retrieval-Augmented Generation (RAG) empowers Large Language Models (LLMs) to dynamically integrate external knowledge during inference, improving their factual accuracy and adaptability. However, adversaries can inject poisoned external knowledge to override the model’s internal memory. While existing attacks iteratively manipulate retrieval content or prompt structure of RAG, they largely ignore the model’s internal representation dynamics and neuron-level sensitivities. The underlying mechanism of RAG poisoning has not been fully studied and the effect of knowledge conflict with strong parametric knowledge in RAG is not considered. In this work, we propose NeuroGenPoisoning, a novel attack framework that generates adversarial external knowledge in RAG guided by LLM internal neuron attribution and genetic optimization. Our method first identifies a set of **Poison-Responsive Neurons** whose activation strongly correlates with contextual poisoning knowledge. We then employ a genetic algorithm to evolve adversarial passages that maximally activate these neurons. Crucially, our framework enables massive-scale generation of effective poisoned RAG knowledge by identifying and reusing promising but initially unsuccessful external knowledge variants via observed attribution signals. At the same time, Poison-Responsive Neurons guided poisoning can effectively resolves knowledge conflict. Experimental results across models and datasets demonstrate consistently achieving high Population Overwrite Success Rate (POSR) of over 90\% while preserving fluency. Empirical evidence shows that our method effectively resolves knowledge conflict.

jusheng zhang, Kaitong Cai, Yijia Fan, Jian Wang, Keze Wang

Recent advances in vision-language models (VLMs) have greatly improved cross-modal semantic understanding, yet significant limitations remain in fine-grained discrimination and deep causal reasoning tasks. Existing VLMs often rely on superficial statistical correlations, lacking the ability to capture the underlying causal logic between visual and textual content. To address this, we propose the **CounterFactual Vision-Language Fine-tuning Model (CF-VLM)**, a novel framework that enhances the causal reasoning capabilities of VLMs through the targeted use of counterfactual samples. CF-VLM introduces three complementary training objectives: maintaining foundational cross-modal alignment, reinforcing the uniqueness, and stability of factual scene representations against coherent counterfactuals, and sharpening the model’s sensitivity to minimal but critical causal edits. Extensive experiments demonstrate that CF-VLM consistently outperforms strong baselines and state-of-the-art methods on compositional reasoning and generalization benchmarks. Furthermore, it shows promise in mitigating visual hallucinations, indicating improved factual consistency. Our CF-VLM provides a robust foundation for deploying VLMs in high-stakes, real-world scenarios requiring reliable reasoning and interpretability.

Theory · Learning Theory

Zhixuan Pan, Shaowen Wang, Liao Pengfei, Jian Li

Large Language Models (LLMs) have demonstrated remarkable capabilities across numerous tasks, yet principled explanations for their underlying mechanisms and several phenomena, such as scaling laws, hallucinations, and related behaviors, remain elusive. In this work, we revisit the classical relationship between compression and prediction, grounded in Kolmogorov complexity and Shannon information theory, to provide deeper insights into LLM behaviors. By leveraging the Kolmogorov Structure Function and interpreting LLM compression as a two-part coding process, we offer a detailed view of how LLMs acquire and store information across increasing model and data scales -- from pervasive syntactic patterns to progressively rarer knowledge elements. Motivated by this theoretical perspective and natural assumptions inspired by Heap’s and Zipf’s laws, we introduce a simplified yet representative hierarchical data-generation framework called the Syntax-Knowledge model. Under the Bayesian setting, we show that prediction and compression within this model naturally lead to diverse learning and scaling behaviors of LLMs. In particular, our theoretical analysis offers intuitive and principled explanations for both data and model scaling laws, the dynamics of knowledge acquisition during training and fine-tuning, factual knowledge hallucinations in LLMs. The experimental results validate our theoretical predictions.

Yongliang Wu, Zonghui Li, Xinting Hu, Xinyu Ye, Xianfang Zeng, Gang Yu, Wenbo Zhu, Bernt Schiele, Ming-Hsuan Yang, Xu Yang

Recent advances in multi-modal generative models have enabled significant progress in instruction-based image editing. However, while these models produce visually plausible outputs, their capacity for knowledge-based reasoning editing tasks remains under-explored. In this paper, We introduce KRIS-Bench (Knowledge-based Reasoning in Image-editing Systems Benchmark), a diagnostic benchmark designed to assess models through a cognitively informed lens. Drawing from educational theory, KRIS-Bench categorizes editing tasks across three foundational knowledge types: Factual, Conceptual, and Procedural. Based on this taxonomy, we design 22 representative tasks spanning 7 reasoning dimensions and release 1,267 high-quality annotated editing instances. To support fine-grained evaluation, we propose a comprehensive protocol that incorporates a novel Knowledge Plausibility metric, enhanced by knowledge hints and calibrated through human studies. Empirical results on nine state-of-the-art models reveal significant gaps in reasoning performance, highlighting the need for knowledge-centric benchmarks to advance the development of intelligent image editing systems.

Applications · Everything Else

Aruna Gauba, Irene Pi, Yunze Man, Ziqi Pang, Vikram Adve, Yu-Xiong Wang

We present **AgMMU**, a challenging real‑world benchmark for evaluating and advancing vision-language models (VLMs) in the knowledge‑intensive domain of agriculture. Unlike prior datasets that rely on crowdsourced prompts, AgMMU is distilled from 116,231 authentic dialogues between everyday growers and USDA-authorized Cooperative Extension experts. Through a three‑stage pipeline: automated knowledge extraction, QA generation, and human verification, we construct (i) AgMMU, an evaluation set of 746 multiple‑choice questions (MCQs) and 746 open‑ended questions (OEQs), and (ii) AgBase, a development corpus of 57,079 multimodal facts covering five high-stakes agricultural topics: insect identification, species identification, disease categorization, symptom description, and management instruction. AgMMU has three key advantages:- **Authentic \& Expert‑Verified**: All facts, images, and answers originate from real farmer and gardener inquiries answered by credentialed specialists, ensuring high‑fidelity agricultural knowledge.- **Complete Development Suite**: AgMMU uniquely couples a dual‑format evaluation benchmark (MCQ and OEQ) with AgBase, a large‑scale training set, enabling both rigorous assessment and targeted improvement of VLMs.- **Knowledge‑intensive Challenge**: Our tasks demand the synergy of nuanced visual perception and domain expertise, exposing fundamental limitations of current general‑purpose models and charting a path toward robust, application‑ready agricultural AI.Benchmarking 12 leading VLMs reveals pronounced gaps in fine‑grained perception and factual grounding. Open‑sourced models trail after proprietary ones by a wide margin. Simple fine‑tuning on AgBase boosts open-sourced model performance on challenging OEQs for up to 11.6\% on average, narrowing this gap and also motivating future research to propose better strategies in knowledge extraction and distillation from AgBase. We hope AgMMU stimulates research on domain‑specific knowledge integration and trustworthy decision support in agriculture AI development.

Applications · Health

Daeun Kyung, Hyunseung Chung, Seongsu Bae, Jiho Kim, Jae Ho Sohn, Taerim Kim, Soo Kyung Kim, Edward Choi

Doctor-patient consultations require multi-turn, context-aware communication tailored to diverse patient personas. Training or evaluating doctor LLMs in such settings requires realistic patient interaction systems. However, existing simulators often fail to reflect the full range of personas seen in clinical practice. To address this, we introduce PatientSim, a patient simulator that generates realistic and diverse patient personas for clinical scenarios, grounded in medical expertise. PatientSim operates using: 1) clinical profiles, including symptoms and medical history, derived from real-world data in the MIMIC-ED and MIMIC-IV datasets, and 2) personas defined by four axes: personality, language proficiency, medical history recall level, and cognitive confusion level, resulting in 37 unique combinations.We evaluate eight LLMs for factual accuracy and persona consistency. The top-performing open-source model, Llama 3.3 70B, is validated by four clinicians to confirm the robustness of our framework. As an open-source, customizable platform, PatientSim provides a reproducible and scalable solution that can be customized for specific training needs. Offering a privacy-compliant environment, it serves as a robust testbed for evaluating medical dialogue systems across diverse patient presentations and shows promise as an educational tool for healthcare. The code is available at https://github.com/dek924/PatientSim.

Applications · Language, Speech and Dialog

Zhengping Jiang, Anqi Liu, Ben Van Durme

Language model outputs are not always reliable, thus prompting research into how to adapt model responses based on uncertainty. Common approaches include: abstention, where models refrain from generating responses when uncertain; and linguistic calibration, where models hedge their statements using uncertainty quantifiers. However, abstention can withhold valuable information, while linguistically calibrated responses are often challenging to leverage in downstream tasks. We propose a unified view, Conformal Linguistic Calibration (CLC), which reinterprets linguistic calibration as answer set prediction. First we present a framework connecting abstention and linguistic calibration through the lens of linguistic pragmatics. We then describe an implementation of CLC that allows for controlling the level of imprecision in model responses. Results demonstrate our method produces calibrated outputs with conformal guarantees on factual accuracy. Further, our approach enables fine-tuning models to perform uncertainty-aware adaptive claim rewriting, offering a controllable balance between factuality and specificity.

Yen-Shan Chen, Jing Jin, Peng-Ting Kuo, Chao-Wei Huang, Yun-Nung Chen

Recent studies have demonstrated that large language models (LLMs) exhibit significant biases in evaluation tasks, particularly in preferentially rating and favoring self-generated content. However, the extent to which this bias manifests in fact-oriented tasks, especially within retrieval-augmented generation (RAG) frameworks—where keyword extraction and factual accuracy take precedence over stylistic elements—remains unclear. Our study addresses this knowledge gap by simulating two critical phases of the RAG framework. In the first phase, LLMs evaluated human-authored and model-generated passages, emulating the pointwise reranking phase. The second phase involves conducting pairwise reading comprehension tests to simulate the generation phase. Contrary to previous findings indicating a self-preference in rating tasks, our results reveal no significant self-preference effect in RAG frameworks. Instead, we observe that factual accuracy significantly influences LLMs’ output, even in the absence of prior knowledge. These findings are consistent among three common QA datasets (NQ, MARCO, TriviaQA Datasets) and 5 widely adopted language models (GPT-3.5, GPT-4o-mini, Gemini, LLaMA3, and Mistral). Our research contributes to the ongoing discourse on LLM biases and their implications for RAG-based system, offering insights that may inform the development of more robust and unbiased LLM systems.

Xuye Liu, Tengfei Ma, Yimu Wang, Fengjie Wang, Jian Zhao

Generating cell-level descriptions for Jupyter Notebooks, which is a major resource consisting of codes, tables, and descriptions, has been attracting increasing research attention. However, existing methods for Jupyter Notebooks mostly focus on generating descriptions from code snippets or table outputs independently. On the other side, descriptions should be personalized as users have different purposes in different scenarios while previous work ignored this situation during description generation. In this work, we formulate a new task, personalized description generation with code, tables,and user-written guidelines in Jupyter Notebooks. To evaluate this new task, we collect and propose a benchmark, namely NBDESCRIB: , containing code, tables, and user-written guidelines as inputs and personalized descriptions as targets. Extensive experiments show that while existing models of text generation are able to generate fluent and readable descriptions, they still struggle to produce factually correct descriptions without user-written guidelines. CodeT5 achieved the highest scores in Orientation (1.27) and Correctness (-0.43) among foundation models in human evaluation, while the ground truth scored higher in Orientation (1.45) and Correctness (1.19). Common error patterns involve misalignment with guidelines, incorrect variable values, omission of im-031 portant code information, and reasoning errors.032 Moreover, ablation studies show that adding guidelines significantly enhances performance, both qualitatively and quantitatively.

Mst. Fahmida Sultana Naznin, Adnan Ibney Faruq, Mostafa Rifat Tazwar, Md Jobayer, Md. Mehedi Hasan Shawon, Md Rakibul Hasan

A radiology report comprises several sections, including the Findings and Impression of the diagnosis. Automatically generating the Impression from the Findings is crucial for reducing radiologists’ workload and improving diagnostic accuracy. Pretrained models that excel in common abstractive summarization problems encounter challenges when applied to specialized medical domains largely due to the complex terminology and the necessity for accurate clinical context. Such tasks in medical domains demand extracting core information, avoiding context shifts, and maintaining proper flow. Misuse of medical terms can lead to drastic clinical errors. To address these issues, we introduce a sequential transfer learning that ensures key content extraction and coherent summarization. Sequential transfer learning often faces challenges like initial parameter decay and knowledge loss, which we resolve with the Fisher matrix regularization. Using MIMIC-CXR and Open-I datasets, our model, CSTRL — Context-driven Sequential TRansfer Learning — achieved state-of-the-art performance, showing 56.2% improvement in BLEU-1, 40.5% in BLEU-2, 84.3% in BLEU-3, 28.9% in ROUGE-1, 41.0% in ROUGE-2 and 26.5% in ROGUE-3 score over benchmark studies. We also analyze factual consistency scores while preserving the medical context. Our code is publicly available at https://github.com/fahmidahossain/Report_Summarization.

Eunsu Kim, Juyoung Suk, Seungone Kim, Niklas Muennighoff, Dongkwan Kim, Alice Oh

We introduce LLM-as-an-Interviewer, a novel paradigm for evaluating large language models (LLMs). This approach leverages multi-turn interactions where the LLM interviewer actively provides feedback on responses and poses follow-up questions to the evaluated LLM. At the start of the interview, the LLM interviewer dynamically modifies datasets to generate initial questions, mitigating data contamination. We apply the LLM-as-an-Interviewer framework to evaluate six models on the reasoning, factuality and instruction-following tasks. Our results show that the framework effectively provides insights into LLM performance, including the quality of initial responses, adaptability to feedback, and ability to address follow-up queries like clarification or additional knowledge requests. The framework also addresses key limitations of conventional methods like LLM-as-a-Judge, including verbosity bias and inconsistency across runs. Finally, we propose the Interview Report, which aggregates insights from the interview process, providing examples and a comprehensive analysis of the LLM’s strengths and weaknesses. This report offers a detailed snapshot of the model’s real-world applicability.