Tenghao Huang, Kung-Hsiang Huang, Prafulla Kumar Choubey, Yilun Zhou, Muhao Chen, Jonathan May, Chien-Sheng Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

To keep pace with the increasing velocity of large language models (LLM) development, model output evaluation has transitioned away from time-consuming human evaluation to automatic evaluation, where LLMs themselves are tasked with assessing and critiquing other model outputs.LLM-as-judge models are a class of generative evaluators that excel in evaluating relatively simple domains, like chat quality, but struggle in reasoning intensive domains where model responses contain more substantive and challenging content.To remedy existing judge shortcomings, we explore training judges with reinforcement learning (RL).We make three key contributions: (1) We propose the Equivalent Initial State Group Relative Policy Optimization (EIS-GRPO) algorithm, which allows us to train our judge to be robust to positional biases that arise in more complex evaluation settings.(2) We introduce ReasoningJudgeBench, a benchmark that evaluates judges in diverse reasoning settings not covered by prior work.(3) We train Judge for Reasoning (J4R), a 7B judge trained with EIS-GRPO that outperforms GPT-4o and the next best small judge by 6.7% and 9%, matching or exceeding the performance of larger GRPOtrained judges on both JudgeBench and Rea-soningJudgeBench.

Large language models (LLMs) demonstrate proficiency across numerous computational tasks, yet their inner workings remain unclear.In theory, the combination of causal self-attention and multilayer perceptron layers allows every token to access and compute information based on all preceding tokens.In practice, to what extent are such operations present?In this paper, on mental math tasks (i.e., direct math calculation via next-token prediction without explicit reasoning), we investigate this question in three steps: inhibiting input-specific token computations in the initial layers, restricting the routes of information transfer across token positions in the next few layers, and forcing all computation to happen at the last token in the remaining layers.With two proposed techniques, Context-Aware Mean Ablation (CAMA) and Attention-Based Peeking (ABP), we identify an All-for-One subgraph (AF1) with high accuracy on a wide variety of mental math tasks, where meaningful computation occurs very late (in terms of layer depth) and only at the last token, which receives information of other tokens in few specific middle layers.Experiments on a variety of models and arithmetic expressions show that this subgraph is sufficient and necessary for high model performance, transfers across different models, and works on a variety of input styles.Ablations on different CAMA and ABP alternatives reveal their unique advantages over other methods, which may be of independent interest.

To meet the increasing need for timely and accurate evaluation of large language model (LLM) responses, training LLM-as-judges to evaluate and critique other model responses has emerged as a popular paradigm.However, existing judge models are largely trained with supervised finetuning (SFT) on small data scales to perform limited types of evaluation tasks, fundamentally limiting generalization.To meet the need for strong, generalized judge models, we explore training foundational judge models at large data scales (680K) with direct preference optimization (DPO).Using four training tasks, we form three types of DPO preference pairs targeting different aspects of evaluation: Generating meaningful critiques, making accurate judgements, and understanding what comprises good and bad responses.To demonstrate the effectiveness of our method, we train judge models of three sizes: 8B parameters, 12B, and 70B, and evaluate on a comprehensive suite of 13 benchmarks (7 pairwise, 4 single rating, and 2 classification).Our models achieve the best aggregate performance, with even our 8B model outperforming GPT-4o in pairwise benchmarks.Further analysis shows that our judge models produce factual and actionable critiques and serve as strong foundational judges for continued finetuning.

Malicious content generated by large language models (LLMs) can pose varying degrees of harm. Although existing LLM-based moderators can detect harmful content, they struggle to assess risk levels and may miss lower-risk outputs. Accurate risk assessment allows platforms with different safety thresholds to tailor content filtering and rejection. In this paper, we introduce per-topic severity rubrics for 11 harmful topics and build BingoGuard, an LLM-based moderation system designed to predict both binary safety labels and severity levels. To address the lack of annotations on levels of severity, we propose a scalable generate-then-filter framework that first generates responses across different severity levels and then filters out low-quality responses. Using this framework, we create BingoGuardTrain, a training dataset with 54,897 examples covering a variety of topics, response severity, styles, and BingoGuardTest, a test set with 988 examples explicitly labeled based on our severity rubrics that enables fine-grained analysis on model behaviors on different severity levels. Our BingoGuard-8B, trained on BingoGuardTrain, achieves the state-of-the-art performance on several moderation benchmarks, including WildGuardTest and HarmBench, as well as BingoGuardTest, outperforming best public models, WildGuard, by 4.3\%. Our analysis demonstrates that incorporating severity levels into training significantly enhances detection performance and enables the model to effectively gauge the severity of harmful responses. Warning: this paper includes red-teaming examples that may be harmful in nature.

Scaling test-time computation, or affording a generator large language model (LLM) extra compute during inference, typically employs the help of external non-generative evaluators (i.e., reward models). Concurrently, LLM-judges, models trained to generate evaluations and critiques (explanations) in natural language, are becoming increasingly popular in automatic evaluation. Despite judge empirical successes, their effectiveness as evaluators in test-time scaling settings is largely unknown. In this paper, we introduce the Judge Evaluation for Test-Time Scaling (JETTS) benchmark, which evaluates judge performance in three domains (math reasoning, code generation, and instruction following) under three task settings: response reranking, step-level beam search, and critique-based response refinement. We evaluate 10 different judge models (7B-70B parameters) for 8 different base generator models (6.7B-72B parameters). Our benchmark shows that while judges are competitive with outcome reward models in reranking, they are consistently worse than process reward models in beam search procedures. Furthermore, though unique to LLM-judges, their natural language critiques are currently ineffective in guiding the generator towards better responses.

While large language models (LLMs) have demonstrated strong capability in structured prediction tasks such as semantic parsing, few amounts of research have explored the underlying mechanisms of their success. Our work studies different methods for explaining an LLM-based semantic parser and qualitatively discusses the explained model behaviors, hoping to inspire future research toward better understanding them.

Counterfactual (CF) explanations, also known as contrastive explanations and algorithmic recourses, are popular for explaining machine learning models in high-stakes domains. For a subject that receives a negative model prediction (e.g., mortgage application denial), the CF explanations are similar instances but with positive predictions, which informs the subject of ways to improve. While their various properties have been studied, such as validity and stability, we contribute a novel one: their behaviors under iterative partial fulfillment (IPF). Specifically, upon receiving a CF explanation, the subject may only partially fulfill it before requesting a new prediction with a new explanation, and repeat until the prediction is positive. Such partial fulfillment could be due to the subject’s limited capability (e.g., can only pay down two out of four credit card accounts at this moment) or an attempt to take the chance (e.g., betting that a monthly salary increase of $800 is enough even though $1,000 is recommended). Does such iterative partial fulfillment increase or decrease the total cost of improvement incurred by the subject? We mathematically formalize IPF and demonstrate, both theoretically and empirically, that different CF algorithms exhibit vastly different behaviors under IPF. We discuss implications of our observations, advocate for this factor to be carefully considered in the development and study of CF algorithms, and give several directions for future work.

Feature attribution methods are popular in interpretable machine learning. These methods compute the attribution of each input feature to represent its importance, but there is no consensus on the definition of "attribution", leading to many competing methods with little systematic evaluation, complicated in particular by the lack of ground truth attribution. To address this, we propose a dataset modification procedure to induce such ground truth. Using this procedure, we evaluate three common methods: saliency maps, rationales, and attentions. We identify several deficiencies and add new perspectives to the growing body of evidence questioning the correctness and reliability of these methods applied on datasets in the wild. We further discuss possible avenues for remedy and recommend new attribution methods to be tested against ground truth before deployment. The code and appendix are available at https://yilunzhou.github.io/feature-attribution-evaluation/.

Interpretability methods are developed to understand the working mechanisms of blackbox models, which is crucial to their responsible deployment.Fulfilling this goal requires both that the explanations generated by these methods are correct and that people can easily and reliably understand them.While the former has been addressed in prior work, the latter is often overlooked, resulting in informal model understanding derived from a handful of local explanations.In this paper, we introduce explanation summary (EXSUM), a mathematical framework for quantifying model understanding, and propose metrics for its quality assessment.On two domains, EXSUM highlights various limitations in the current practice, helps develop accurate model understanding, and reveals easily overlooked properties of the model.We also connect understandability to other properties of explanations such as human alignment, robustness, and counterfactual similarity and plausibility.

We envision a world in which robots serve as capable partners in heterogeneous teams composed of other robots or humans. A crucial step towards such a world is enabling robots to learn to use the same representations as their partners; with a shared representation scheme, information may be passed among teammates. We define the problem of learning a fixed partner’s representation scheme as that of latent space alignment and propose metrics for evaluating the quality of alignment. While techniques from prior art in other fields may be applied to the latent space alignment problem, they often require interaction with partners during training time or large amounts of training data. We developed a technique, Adversarially Guided Self-Play (ASP), that trains agents to solve the latent space alignment problem with little training data and no access to their pre-trained partners. Simulation results confirmed that, despite using less training data, agents trained by ASP aligned better with other agents than agents trained by other techniques. Subsequent human-participant studies involving hundreds of Amazon Mechanical Turk workers showed how laypeople understood our machines enough to perform well on team tasks and anticipate their machine partner’s successes or failures.

Post-hoc explanation methods are gaining popularity for interpreting, understanding, and debugging neural networks. Most analyses using such methods explain decisions in response to inputs drawn from the test set. However, the test set may have few examples that trigger some model behaviors, such as high-confidence failures or ambiguous classifications. To address these challenges, we introduce a flexible model inspection framework: Bayes-TrEx. Given a data distribution, Bayes-TrEx finds in-distribution examples which trigger a specified prediction confidence. We demonstrate several use cases of Bayes-TrEx, including revealing highly confident (mis)classifications, visualizing class boundaries via ambiguous examples, understanding novel-class extrapolation behavior, and exposing neural network overconfidence. We use Bayes-TrEx to study classifiers trained on CLEVR, MNIST, and Fashion-MNIST, and we show that this framework enables more flexible holistic model analysis than just inspecting the test set. Code and supplemental material are available at https://github.com/serenabooth/Bayes-TrEx.

Active learning (AL) algorithms may achieve better performance with fewer data because the model guides the data selection process. While many algorithms have been proposed, there is little study on what the optimal AL algorithm looks like, which would help researchers understand where their models fall short and iterate on the design. In this paper, we present a simulated annealing algorithm to search for this optimal oracle and analyze it for several tasks. We present qualitative and quantitative insights into the behaviors of this oracle, comparing and contrasting them with those of various heuristics. Moreover, we are able to consistently improve the heuristics using one particular insight. We hope that our findings can better inform future active learning research. The code is available at https://github.com/YilunZhou/optimal-active-learning.

In many applications, it is important to characterize the way in which two concepts are semantically related. Knowledge graphs such as ConceptNet provide a rich source of information for such characterizations by encoding relations between concepts as edges in a graph. When two concepts are not directly connected by an edge, their relationship can still be described in terms of the paths that connect them. Unfortunately, many of these paths are uninformative and noisy, which means that the success of applications that use such path features crucially relies on their ability to select high-quality paths. In existing applications, this path selection process is based on relatively simple heuristics. In this paper we instead propose to learn to predict path quality from crowdsourced human assessments. Since we are interested in a generic task-independent notion of quality, we simply ask human participants to rank paths according to their subjective assessment of the paths' naturalness, without attempting to define naturalness or steering the participants towards particular indicators of quality. We show that a neural network model trained on these assessments is able to predict human judgments on unseen paths with near optimal performance. Most notably, we find that the resulting path selection method is substantially better than the current heuristic approaches at identifying meaningful paths.

Convolutional neural networks (CNN) are a deep learning technique that has achieved state-of-the-art prediction performance in computer vision and robotics, but assume the input data can be formatted as an image or video (e.g. predicting a robot grasping location given RGB-D image input). This paper considers the problem of augmenting a traditional CNN for handling image-like input (called main-channel input) with additional, highly predictive, non-image-like input (called side-channel input). An example of such a task would be to predict whether a robot path is collision-free given an occupancy grid of the environment and the path's start and goal configurations; the occupancy grid is the main-channel and the start and goal are the side-channel. This paper presents several candidate network architectures for doing so. Empirical tests on robot collision prediction and control problems compare the proposed architectures in terms of learning speed, memory usage, learning capacity, and susceptibility to overfitting.

This paper presents an equivalence between feasible kinodynamic planning and optimal kinodynamic planning, in that any optimal planning problem can be transformed into a series of feasible planning problems in a state-cost space, whose solutions approach the optimum. This transformation yields a meta-algorithm that produces an asymptotically optimal planner, given any feasible kinodynamic planner as a subroutine. The meta-algorithm is proven to be asymptotically optimal and a formula is derived relating expected running time and solution suboptimality. It is directly applicable to a wide range of optimal planning problems because it does not resort to the use of steering functions or numerical boundary-value problem solvers. On a set of benchmark problems, it is demonstrated to perform, using the expansive space tree (EST) and rapidly-exploring random tree (RRT) algorithms as subroutines, at a level that is superior or comparable to related planners.