While the Question Generation (QG) task has been increasingly adopted in educational assessments, its evaluation remains limited by approaches that lack a clear connection to the educational values of test items. In this work, we introduce test item analysis, a method frequently used by educators to assess test question quality, into QG evaluation. Specifically, we construct pairs of candidate questions that differ in quality across dimensions such as topic coverage, item difficulty, item discrimination, and distractor efficiency. We then examine whether existing QG evaluation approaches can effectively distinguish these differences. Our findings reveal significant shortcomings in these approaches with respect to accurately assessing test item quality in relation to student performance. To address this gap, we propose a novel QG evaluation framework, QG-SMS, which leverages Large Language Model for Student Modeling and Simulation to perform test item analysis. As demonstrated in our extensive experiments and human evaluation study, the additional perspectives introduced by the simulated student profiles lead to a more effective and robust assessment of test items.
A large survey of American adults explored the complex landscape of attitudes towards artificial intelligence (AI). It explored the degree of concern regarding specific potential outcomes of the new advances in AI technology and correlates of these concerns. Key variables associated with the direction and intensity of concern include prior experience using a large language model such as ChatGPT, general trust in science, adherence to the precautionary principle versus support for unrestricted innovation, and demographic factors such as gender. By analyzing these relationships, the paper provides valuable insights into the American public's response to AI that are particularly important in the development of policy to regulate or further encourage its development.
Retrieval-augmented generation (RAG) enhances the reliability of large language model (LLM) answers by integrating external knowledge. However, RAG increases the end-to-end inference time since looking for relevant documents from large vector databases is computationally expensive. To address this, we introduce Proximity, an approximate key-value cache that optimizes the RAG workflow by leveraging similarities in user queries. Instead of treating each query independently, Proximity reuses previously retrieved documents when similar queries appear, reducing reliance on expensive vector database lookups. We evaluate Proximity on the MMLU and MedRAG benchmarks, demonstrating that it significantly improves retrieval efficiency while maintaining response accuracy. Proximity reduces retrieval latency by up to 59% while maintaining accuracy and lowers the computational burden on the vector database. We also experiment with different similarity thresholds and quantify the trade-off between speed and recall. Our work shows that approximate caching is a viable and effective strategy for optimizing RAG-based systems.
This paper investigates a critical aspect of large language model (LLM) performance: the optimal formatting of classification task options in prompts. Through an extensive experimental study, we compared two selection formats -- bullet points and plain English -- to determine their impact on model performance. Our findings suggest that presenting options via bullet points generally yields better results, although there are some exceptions. Furthermore, our research highlights the need for continued exploration of option formatting to drive further improvements in model performance.
Large Language Model (LLM) based judges form the underpinnings of key safety evaluation processes such as offline benchmarking, automated red-teaming, and online guardrailing. This widespread requirement raises the crucial question: can we trust the evaluations of these evaluators? In this paper, we highlight two critical challenges that are typically overlooked: (i) evaluations in the wild where factors like prompt sensitivity and distribution shifts can affect performance and (ii) adversarial attacks that target the judge. We highlight the importance of these through a study of commonly used safety judges, showing that small changes such as the style of the model output can lead to jumps of up to 0.24 in the false negative rate on the same dataset, whereas adversarial attacks on the model generation can fool some judges into misclassifying 100% of harmful generations as safe ones. These findings reveal gaps in commonly used meta-evaluation benchmarks and weaknesses in the robustness of current LLM judges, indicating that low attack success under certain judges could create a false sense of security.
As large language models (LLMs) become increasingly capable of autonomous decision-making, they introduce new challenges and opportunities for human-AI cooperation in mixed-motive contexts. While prior research has primarily examined AI in assistive or cooperative roles, little is known about how humans interact with AI agents perceived as independent and strategic actors. This study investigates human cooperative attitudes and behaviors toward LLM agents by engaging 30 participants (15 males, 15 females) in repeated Prisoner's Dilemma games with agents differing in declared identity: purported human, rule-based AI, and LLM agent. Behavioral metrics, including cooperation rate, decision latency, unsolicited cooperative acts and trust restoration tolerance, were analyzed to assess the influence of agent identity and participant gender. Results revealed significant effects of declared agent identity on most cooperation-related behaviors, along with notable gender differences in decision latency. Furthermore, qualitative responses suggest that these behavioral differences were shaped by participants interpretations and expectations of the agents. These findings contribute to our understanding of human adaptation in competitive cooperation with autonomous agents and underscore the importance of agent framing in shaping effective and ethical human-AI interaction.
We explore the political and ideological positioning of ChatGPT, a leading large language model (LLM), by comparing its responses to political economy questions from the European Social Survey (ESS). The questions concern environmental sustainability, civil rights, income inequality, and government size. ChatGPT's self-assessed placement on a left-right political spectrum is compared to the ideological stances of individuals providing similar answers in the ESS dataset. Results highlight a significant left-oriented bias in ChatGPT's answers, particularly on environmental and civil rights topics, diverging from its same self-declared center-left stance. These findings underscore the need for transparency in AI systems to prevent potential ideological influences on users. We conclude by discussing the implications for AI governance, debiasing strategies, and educational use.
Large language models (LLMs) are being used in many applications and prompts for these models are integrated into software applications as code-like artifacts. These prompts behave much like traditional software in that they take inputs, generate outputs, and perform some specific function. However, prompts differ from traditional code in many ways and require new approaches to ensure that they are robust. For example, unlike traditional software the output of a prompt depends on the AI model that interprets it. Also, while natural language prompts are easy to modify, the impact of updates is harder to predict. New approaches to testing, debugging, and modifying prompts with respect to the model running them are required. To address some of these issues, we developed PromptPex, an LLM-based tool to automatically generate and evaluate unit tests for a given prompt. PromptPex extracts input and output specifications from a prompt and uses them to generate diverse, targeted, and valid unit tests. These tests are instrumental in identifying regressions when a prompt is changed and also serve as a tool to understand how prompts are interpreted by different models. We use PromptPex to generate tests for eight benchmark prompts and evaluate the quality of the generated tests by seeing if they can cause each of four diverse models to produce invalid output. PromptPex consistently creates tests that result in more invalid model outputs than a carefully constructed baseline LLM-based test generator. Furthermore, by extracting concrete specifications from the input prompt, PromptPex allows prompt writers to clearly understand and test specific aspects of their prompts. The source code of PromptPex is available at https://github.com/microsoft/promptpex.
Decoder-based transformers, while revolutionizing language modeling and scaling to immense sizes, have not completely overtaken encoder-heavy architectures in natural language processing. Specifically, encoder-only models remain dominant in tasks like classification, regression, and ranking. This is primarily due to the inherent structure of decoder-based models, which limits their direct applicability to these tasks. In this paper, we introduce Gemma Encoder, adapting the powerful Gemma decoder model to an encoder architecture, thereby unlocking its potential for a wider range of non-generative applications. To optimize the adaptation from decoder to encoder, we systematically analyze various pooling strategies, attention mechanisms, and hyperparameters (e.g., dropout rate). Furthermore, we benchmark Gemma Encoder against established approaches on the GLUE benchmarks, and MS MARCO ranking benchmark, demonstrating its effectiveness and versatility.
Collecting ground truth task completion rewards or human demonstrations for multi-step reasoning tasks is often cost-prohibitive and time-consuming, especially in interactive domains like web tasks. To address this bottleneck, we present self-taught lookahead, a self-supervised method that leverages state-transition dynamics to train a value model capable of effectively guiding language model-controlled search. We find that moderately sized (8 billion parameters) open-weight value models improved with self-taught lookahead can match the performance of using a frontier LLM such as gpt-4o as the value model. Furthermore, we find that self-taught lookahead improves performance by 20% while reducing costs 37x compared to previous LLM-based tree search, without relying on ground truth rewards.
In modern large language models (LLMs), LLM alignment is of crucial importance and is typically achieved through methods such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). However, in most existing methods for LLM alignment, all tokens in the response are optimized using a sparse, response-level reward or preference annotation. The ignorance of token-level rewards may erroneously punish high-quality tokens or encourage low-quality tokens, resulting in suboptimal performance and slow convergence speed. To address this issue, we propose AlignDistil, an RLHF-equivalent distillation method for token-level reward optimization. Specifically, we introduce the reward learned by DPO into the RLHF objective and theoretically prove the equivalence between this objective and a token-level distillation process, where the teacher distribution linearly combines the logits from the DPO model and a reference model. On this basis, we further bridge the accuracy gap between the reward from the DPO model and the pure reward model, by building a contrastive DPO reward with a normal and a reverse DPO model. Moreover, to avoid under- and over-optimization on different tokens, we design a token adaptive logit extrapolation mechanism to construct an appropriate teacher distribution for each token. Experimental results demonstrate the superiority of our AlignDistil over existing methods and showcase fast convergence due to its token-level distributional reward optimization.
An essential component of modern recurrent sequence models is the forget gate. While Transformers do not have an explicit recurrent form, we show that a forget gate can be naturally incorporated into Transformers by down-weighting the unnormalized attention scores in a data-dependent way. We name this attention mechanism Forgetting Attention and the resulting model the Forgetting Transformer (FoX). We show that FoX outperforms the Transformer on long-context language modeling, length extrapolation, and short-context downstream tasks, while performing on par with the Transformer on long-context downstream tasks. Moreover, it is compatible with the FlashAttention algorithm and does not require any positional embeddings. Several analyses, including the needle-in-the-haystack test, show that FoX also retains the Transformer's superior long-context capabilities over recurrent sequence models such as Mamba-2, HGRN2, and DeltaNet. We also introduce a "Pro" block design that incorporates some common architectural components in recurrent sequence models and find it significantly improves the performance of both FoX and the Transformer. Our code is available at https://github.com/zhixuan-lin/forgetting-transformer.
Face manipulation techniques have achieved significant advances, presenting serious challenges to security and social trust. Recent works demonstrate that leveraging multimodal models can enhance the generalization and interpretability of face forgery detection. However, existing annotation approaches, whether through human labeling or direct Multimodal Large Language Model (MLLM) generation, often suffer from hallucination issues, leading to inaccurate text descriptions, especially for high-quality forgeries. To address this, we propose Face Forgery Text Generator (FFTG), a novel annotation pipeline that generates accurate text descriptions by leveraging forgery masks for initial region and type identification, followed by a comprehensive prompting strategy to guide MLLMs in reducing hallucination. We validate our approach through fine-tuning both CLIP with a three-branch training framework combining unimodal and multimodal objectives, and MLLMs with our structured annotations. Experimental results demonstrate that our method not only achieves more accurate annotations with higher region identification accuracy, but also leads to improvements in model performance across various forgery detection benchmarks. Our Codes are available in https://github.com/skJack/VLFFD.git.
Research in linguistics shows that non-verbal cues, such as gestures, play a crucial role in spoken discourse. For example, speakers perform hand gestures to indicate topic shifts, helping listeners identify transitions in discourse. In this work, we investigate whether the joint modeling of gestures using human motion sequences and language can improve spoken discourse modeling in language models. To integrate gestures into language models, we first encode 3D human motion sequences into discrete gesture tokens using a VQ-VAE. These gesture token embeddings are then aligned with text embeddings through feature alignment, mapping them into the text embedding space. To evaluate the gesture-aligned language model on spoken discourse, we construct text infilling tasks targeting three key discourse cues grounded in linguistic research: discourse connectives, stance markers, and quantifiers. Results show that incorporating gestures enhances marker prediction accuracy across the three tasks, highlighting the complementary information that gestures can offer in modeling spoken discourse. We view this work as an initial step toward leveraging non-verbal cues to advance spoken language modeling in language models.
In this work, we benchmark various graph-based retrieval-augmented generation (RAG) systems across a broad spectrum of query types, including OLTP-style (fact-based) and OLAP-style (thematic) queries, to address the complex demands of open-domain question answering (QA). Traditional RAG methods often fall short in handling nuanced, multi-document synthesis tasks. By structuring knowledge as graphs, we can facilitate the retrieval of context that captures greater semantic depth and enhances language model operations. We explore graph-based RAG methodologies and introduce TREX, a novel, cost-effective alternative that combines graph-based and vector-based retrieval techniques. Our benchmarking across four diverse datasets highlights the strengths of different RAG methodologies, demonstrates TREX's ability to handle multiple open-domain QA types, and reveals the limitations of current evaluation methods. In a real-world technical support case study, we demonstrate how TREX solutions can surpass conventional vector-based RAG in efficiently synthesizing data from heterogeneous sources. Our findings underscore the potential of augmenting large language models with advanced retrieval and orchestration capabilities, advancing scalable, graph-based AI solutions.
Large language model (LLM) agents typically adopt a step-by-step reasoning framework, in which they interleave the processes of thinking and acting to accomplish the given task. However, this paradigm faces a deep-rooted one-pass issue whereby each generated intermediate thought is plugged into the trajectory regardless of its correctness, which can cause irreversible error propagation. To address the issue, this paper proposes a novel framework called Generator-Assistant Stepwise Rollback (GA-Rollback) to induce better decision-making for LLM agents. Particularly, GA-Rollback utilizes a generator to interact with the environment and an assistant to examine each action produced by the generator, where the assistant triggers a rollback operation upon detection of incorrect actions. Moreover, we introduce two additional strategies tailored for the rollback scenario to further improve its effectiveness. Extensive experiments show that GA-Rollback achieves significant improvements over several strong baselines on three widely used benchmarks. Our analysis further reveals that GA-Rollback can function as a robust plug-and-play module, integrating seamlessly with other methods.
While Large Language Model-based agents have demonstrated substantial progress in task completion, existing evaluation benchmarks tend to overemphasize single-task performance, with insufficient attention given to the crucial aspects of multitask planning and execution efficiency required in real-world scenarios. To bridge this gap, we present Recipe2Plan, a novel benchmark framework based on real-world cooking scenarios. Unlike conventional benchmarks, Recipe2Plan challenges agents to optimize cooking time through parallel task execution while respecting temporal constraints i.e. specific actions need to be performed within a particular time intervals following the preceding steps. Overly aggressive local parallelization may disrupt this constraint, potentially compromising the entire cooking process. This strict time constraint between actions raises a unique challenge for agents to balance between maximizing concurrent operations and adhering to critical timing constraints. Extensive experiments with state-of-the-art models reveal challenges in maintaining this balance between efficiency and feasibility. The results highlight the need for improved temporal awareness and global multitasking capabilities in large language models. We open-source our benchmark and code at https://github.com/WilliamZR/Recipe2Plan.
Large Language Model (LLM) agents have demonstrated remarkable generalization capabilities across multi-domain tasks. Existing agent tuning approaches typically employ supervised finetuning on entire expert trajectories. However, behavior-cloning of full trajectories can introduce expert bias and weaken generalization to states not covered by the expert data. Additionally, critical steps, such as planning, complex reasoning for intermediate subtasks, and strategic decision-making, are essential to success in agent tasks, so learning these steps is the key to improving LLM agents. For more effective and efficient agent tuning, we propose ATLaS that identifies the critical steps in expert trajectories and finetunes LLMs solely on these steps with reduced costs. By steering the training's focus to a few critical steps, our method mitigates the risk of overfitting entire trajectories and promotes generalization across different environments and tasks. In extensive experiments, an LLM finetuned on only 30% critical steps selected by ATLaS outperforms the LLM finetuned on all steps and recent open-source LLM agents. ATLaS maintains and improves base LLM skills as generalist agents interacting with diverse environments.
Language model evaluations often fail to characterize consequential failure modes, forcing experts to inspect outputs and build new benchmarks. We introduce task elicitation, a method that automatically builds new evaluations to profile model behavior. Task elicitation finds hundreds of natural-language tasks -- an order of magnitude more than prior work -- where frontier models exhibit systematic failures, in domains ranging from forecasting to online harassment. For example, we find that Sonnet 3.5 over-associates quantum computing and AGI and that o3-mini is prone to hallucination when fabrications are repeated in-context.
We assess the vulnerability of multimodal large language models to misleading visualizations - charts that distort the underlying data using techniques such as truncated or inverted axes, leading readers to draw inaccurate conclusions that may support misinformation or conspiracy theories. Our analysis shows that these distortions severely harm multimodal large language models, reducing their question-answering accuracy to the level of the random baseline. To mitigate this vulnerability, we introduce six inference-time methods to improve performance of MLLMs on misleading visualizations while preserving their accuracy on non-misleading ones. The most effective approach involves (1) extracting the underlying data table and (2) using a text-only large language model to answer questions based on the table. This method improves performance on misleading visualizations by 15.4 to 19.6 percentage points.