Ben Zhou

Xingyu Fu, Ben Zhou, Ishaan Preetam Chandratreya et al.

Images are often more significant than only the pixels to human eyes, as we can infer, associate, and reason with contextual information from other sources to establish a more complete picture. For example, in Figure 1, we can find a way to identify the news articles related to the picture through segment-wise understandings of the signs, the buildings, the crowds, and more. This reasoning could provide the time and place the image was taken, which will help us in subsequent tasks, such as automatic storyline construction, correction of image source in intended effect photographs, and upper-stream processing such as image clustering for certain location or time. In this work, we formulate this problem and introduce TARA: a dataset with 16k images with their associated news, time, and location, automatically extracted from New York Times, and an additional 61k examples as distant supervision from WIT. On top of the extractions, we present a crowdsourced subset in which we believe it is possible to find the images' spatio-temporal information for evaluation purpose. We show that there exists a $70\%$ gap between a state-of-the-art joint model and human performance, which is slightly filled by our proposed model that uses segment-wise reasoning, motivating higher-level vision-language joint models that can conduct open-ended reasoning with world knowledge. The data and code are publicly available at https://github.com/zeyofu/TARA.

Nan Xu, Fei Wang, Ben Zhou et al.

While large language models (LLMs) have demonstrated increasing power, they have also given rise to a wide range of harmful behaviors. As representatives, jailbreak attacks can provoke harmful or unethical responses from LLMs, even after safety alignment. In this paper, we investigate a novel category of jailbreak attacks specifically designed to target the cognitive structure and processes of LLMs. Specifically, we analyze the safety vulnerability of LLMs in the face of (1) multilingual cognitive overload, (2) veiled expression, and (3) effect-to-cause reasoning. Different from previous jailbreak attacks, our proposed cognitive overload is a black-box attack with no need for knowledge of model architecture or access to model weights. Experiments conducted on AdvBench and MasterKey reveal that various LLMs, including both popular open-source model Llama 2 and the proprietary model ChatGPT, can be compromised through cognitive overload. Motivated by cognitive psychology work on managing cognitive load, we further investigate defending cognitive overload attack from two perspectives. Empirical studies show that our cognitive overload from three perspectives can jailbreak all studied LLMs successfully, while existing defense strategies can hardly mitigate the caused malicious uses effectively.

Ben Zhou, Dian Yu, Dong Yu et al. · tencent-ai

Speaker identification, determining which character said each utterance in literary text, benefits many downstream tasks. Most existing approaches use expert-defined rules or rule-based features to directly approach this task, but these approaches come with significant drawbacks, such as lack of contextual reasoning and poor cross-lingual generalization. In this work, we propose a speaker identification framework that addresses these issues. We first extract large-scale distant supervision signals in English via general-purpose tools and heuristics, and then apply these weakly-labeled instances with a focus on encouraging contextual reasoning to train a cross-lingual language model. We show that the resulting model outperforms previous state-of-the-art methods on two English speaker identification benchmarks by up to 9% in accuracy and 5% with only distant supervision, as well as two Chinese speaker identification datasets by up to 4.7%.

Ben Zhou, Kyle Richardson, Xiaodong Yu et al.

Explicit decomposition modeling, which involves breaking down complex tasks into more straightforward and often more interpretable sub-tasks, has long been a central theme in developing robust and interpretable NLU systems. However, despite the many datasets and resources built as part of this effort, the majority have small-scale annotations and limited scope, which is insufficient to solve general decomposition tasks. In this paper, we look at large-scale intermediate pre-training of decomposition-based transformers using distant supervision from comparable texts, particularly large-scale parallel news. We show that with such intermediate pre-training, developing robust decomposition-based models for a diverse range of tasks becomes more feasible. For example, on semantic parsing, our model, DecompT5, improves 20% to 30% on two datasets, Overnight and TORQUE, over the baseline language model. We further use DecompT5 to build a novel decomposition-based QA system named DecompEntail, improving over state-of-the-art models, including GPT-3, on both HotpotQA and StrategyQA by 8% and 4%, respectively.

Yu Feng, Ben Zhou, Haoyu Wang et al.

Temporal reasoning is the task of predicting temporal relations of event pairs. While temporal reasoning models can perform reasonably well on in-domain benchmarks, we have little idea of these systems' generalizability due to existing datasets' limitations. In this work, we introduce a novel task named TODAY that bridges this gap with temporal differential analysis, which as the name suggests, evaluates whether systems can correctly understand the effect of incremental changes. Specifically, TODAY introduces slight contextual changes for given event pairs, and systems are asked to tell how this subtle contextual change would affect relevant temporal relation distributions. To facilitate learning, TODAY also annotates human explanations. We show that existing models, including GPT-3.5, drop to random guessing on TODAY, suggesting that they heavily rely on spurious information rather than proper reasoning for temporal predictions. On the other hand, we show that TODAY's supervision style and explanation annotations can be used in joint learning, encouraging models to use more appropriate signals during training and thus outperform across several benchmarks. TODAY can also be used to train models to solicit incidental supervision from noisy sources such as GPT-3.5, thus moving us more toward the goal of generic temporal reasoning systems.

Jiayi Zhang, Jianing Yin, Ben Zhou et al.

LLMs can plan by either generating action sequences directly as a Planner or translating tasks into domain specific language for an external solver as a Formalizer. While most real-world tasks are asynchronous with non-uniform durations, concurrency, and execution-time constraints, existing benchmarks hardly cover them. We unify these asynchronous planning challenges under a single formulation and introduce the first three benchmarks that address each at scale. We conclude that the choice of formal representation primarily determines whether planning scales: as dependency graphs grow from 5 to 100 actions, Planner collapses from 96% to 5% plan accuracy and PDDL2.1 Formalizer from 13% to 0%, while CP-SAT Formalizer averages 94% and still achieves 83% at 100 actions. Faithfulness diagnostics show that PDDL2.1's predicate-based planning representation becomes brittle compared to general constraint satisfaction programs, when LLMs must keep predicates, effects, and goals consistent. Execution-time updates of planning constraints further degrade performance sharply (Planner 23.9%, PDDL2.1 0.7%, CP-SAT 46.1%), but a state-aware repair strategy that updates only event-induced constraints recovers CP-SAT Formalizer to 84.5%.

Zhaonan Li, Kyle R. Chickering, Bangzheng Li et al.

A useful test of visual concept learning is not just whether a model can recognize a concept in a single image, but whether it can preserve and manipulate concept-level properties under transformation and transfer them to new scenes. We introduce VisAnalog, a controlled suite for this setting on natural images. Each example instantiates $A\!:\!B::C\!:\,?$: images $B$ and a hidden target image $D$ are produced by applying the same deterministic transformation sequence to source images $A$ and $C$. Given $A$, $B$, and $C$, a model must answer a multiple-choice question about $D$. The benchmark contains 617 human-validated questions spanning one- to four-step transformations such as zoom, quadrant swap, rotation, flip, and hue rotation. Across strong proprietary and open-source VLMs, end-to-end accuracy is substantially lower than oracle accuracy when $D$ is directly shown, and degrades sharply as transformation depth increases, while human performance remains near the ceiling. A program-conditioned evaluation further separates failures of relation inference from failures of transformation application, showing that inferring the visual relation from $A \rightarrow B$ is the dominant bottleneck, with additional application errors emerging on harder multi-step cases. The dataset is publicly available at https://huggingface.co/datasets/zli99/VisAnalog.

Bangzheng Li, Ben Zhou, Fei Wang et al.

Despite the recent advancement in large language models (LLMs) and their high performances across numerous benchmarks, recent research has unveiled that LLMs suffer from hallucinations and unfaithful reasoning. This work studies a specific type of hallucination induced by semantic associations. Specifically, we investigate to what extent LLMs take shortcuts from certain keyword/entity biases in the prompt instead of following the correct reasoning path. To quantify this phenomenon, we propose a novel probing method and benchmark called EureQA. We start from questions that LLMs will answer correctly with utmost certainty, and mask the important entity with evidence sentence recursively, asking models to find masked entities according to a chain of evidence before answering the question. During the construction of the evidence, we purposefully replace semantic clues (entities) that may lead to the correct answer with distractor clues (evidence) that will not directly lead to the correct answer but require a chain-like reasoning process. We evaluate if models can follow the correct reasoning chain instead of short-cutting through distractor clues. We find that existing LLMs lack the necessary capabilities to follow correct reasoning paths and resist the attempt of greedy shortcuts. We show that the distractor semantic associations often lead to model hallucination, which is strong evidence that questions the validity of current LLM reasoning.

Xiaodong Yu, Ben Zhou, Dan Roth

Information retrieval (IR) or knowledge retrieval, is a critical component for many down-stream tasks such as open-domain question answering (QA). It is also very challenging, as it requires succinctness, completeness, and correctness. In recent works, dense retrieval models have achieved state-of-the-art (SOTA) performance on in-domain IR and QA benchmarks by representing queries and knowledge passages with dense vectors and learning the lexical and semantic similarity. However, using single dense vectors and end-to-end supervision are not always optimal because queries may require attention to multiple aspects and event implicit knowledge. In this work, we propose an information retrieval pipeline that uses entity/event linking model and query decomposition model to focus more accurately on different information units of the query. We show that, while being more interpretable and reliable, our proposed pipeline significantly improves passage coverages and denotation accuracies across five IR and QA benchmarks. It will be the go-to system to use for applications that need to perform IR on a new domain without much dedicated effort, because of its superior interpretability and cross-domain performance.

Zhikun Xu, Yu Feng, Jacob Dineen et al.

Large language model agents trained with reinforcement learning (RL) often learn brittle, task-specific shortcuts. We hypothesize that agents generalize better when their successful trajectories are structurally compressible, decomposed into a small set of reusable abstract patterns. To formalize this, we introduce ReuseRL, which grounds agentic RL in the Minimum Description Length (MDL) principle. ReuseRL extracts a shared skill dictionary from successful trajectories and augments the RL objective with a segmentation cost, explicitly penalizing idiosyncratic behaviors that encode poorly. We prove a PAC-Bayes generalization bound for this compression penalty. Across ALFWorld, TextWorld-Cooking, and Countdown-Stepwise, ReuseRL improves in- and out-of-distribution success over vanilla GRPO and strong round-length baselines.

Dongwon Jung, Qin Liu, Tenghao Huang et al.

Retrieval-augmented generation (RAG) improves large language models (LMs) by incorporating non-parametric knowledge through evidence retrieved from external sources. However, it often struggles to cope with inconsistent and irrelevant information that can distract the LM from its tasks, especially when multiple evidence pieces are required. While compressing the retrieved evidence with a compression model aims to address this issue, the compressed evidence may still be unfamiliar to the target model used for downstream tasks, potentially failing to utilize the evidence effectively. We propose FaviComp (Familarity-Aware Evidence Compression), a novel training-free evidence compression technique that makes retrieved evidence more familiar to the target model, while seamlessly integrating parametric knowledge from the model. Experimental results show that FaviComp consistently outperforms most recent evidence compression baselines across multiple open-domain QA datasets, improving accuracy by up to 28.1% while achieving high compression rates. Additionally, we demonstrate the effective integration of both parametric and non-parametric knowledge during evidence compression.

Sihao Chen, Hongming Zhang, Tong Chen et al.

We introduce sub-sentence encoder, a contrastively-learned contextual embedding model for fine-grained semantic representation of text. In contrast to the standard practice with sentence embeddings, where the meaning of an entire sequence of text is encoded into a fixed-length vector, the sub-sentence encoder learns to produce distinct contextual embeddings corresponding to different atomic propositions, i.e. atomic units of meaning expressed within a text sequence. The sub-sentence embeddings are contrastively learned to recognize (inferred) semantic equivalence between propositions across different text sequences. Our experiments show the effectiveness of sub-sentence encoders in applications, such as retrieving supporting facts for fine-grained text attribution or recognizing the conditional semantic similarity between texts. In practice, we demonstrate that sub-sentence encoders keep the same level of inference cost and space complexity compared to sentence encoders.

Aswin RRV, Jacob Dineen, Divij Handa et al.

Recent advances in test-time scaling have led to the emergence of thinking LLMs that exhibit self-reflective behaviors and multi-step reasoning. While RL drives this self-improvement paradigm, a recent study (Gandhi et al., 2025) shows that RL alone does not truly instill these new reasoning abilities - it merely draws out behaviors already present in the base models. This raises a question: How can we train the models that don't exhibit such thinking behavior to develop it in the first place? To this end, we propose ThinkTuning, a GRPO-based interactive training approach where we augment the rollouts of a student model with the guidance from a teacher model. A simple idea from classroom practice inspires our method: a teacher poses a problem, lets the student try an answer, then gives corrective feedback -- enough to point the mind in the right direction and then show the solution. Each piece of feedback reshapes the student's thoughts, leading them to arrive at the correct solution. Similarly, we find that this type of implicit supervision through feedback from a teacher model of the same size improves the reasoning capabilities of the student model. In particular, on average, our method shows a 3.85% improvement over zero-shot baselines across benchmarks, and on MATH-500, AIME and GPQA-Diamond it shows 2.08%, 2.23% and 3.99% improvements over the vanilla-GRPO baseline. Source code is available at https://github.com/3rdAT/ThinkTuning.

Zhaonan Li, Shijie Lu, Fei Wang et al.

End-to-end Vision-language Models (VLMs) often answer visual questions by exploiting spurious correlations instead of causal visual evidence, and can become more shortcut-prone when fine-tuned. We introduce VISTA (Visual-Information Separation for Text-based Analysis), a modular framework that decouples perception from reasoning via an explicit information bottleneck. A frozen VLM sensor is restricted to short, objective perception queries, while a text-only LLM reasoner decomposes each question, plans queries, and aggregates visual facts in natural language. This controlled interface defines a reward-aligned environment for training unbiased visual reasoning with reinforcement learning. Instantiated with Qwen2.5-VL and Llama3.2-Vision sensors, and trained with GRPO from only 641 curated multi-step questions, VISTA significantly improves robustness to real-world spurious correlations on SpuriVerse (+16.29% with Qwen-2.5-VL-7B and +6.77% with Llama-3.2-Vision-11B), while remaining competitive on MMVP and a balanced SeedBench subset. VISTA transfers robustly across unseen VLM sensors and is able to recognize and recover from VLM perception failures. Human analysis further shows that VISTA's reasoning traces are more neutral, less reliant on spurious attributes, and more explicitly grounded in visual evidence than end-to-end VLM baselines.

Jacob Dineen, Aswin RRV, Zhikun Xu et al.

Co-evolutionary self-play, where one language model generates problems and another solves them, promises autonomous curriculum learning without human supervision. In practice, the proposer quickly converges to a narrow distribution of problems that satisfy the reward function. This diversity collapse renders the curriculum uninformative for the solver, stalling the co-evolutionary loop. We introduce vocabulary dropout, a random mask applied to the proposer's output logits during both policy training and curriculum generation, as a lightweight mechanism to sustain diversity. The mask is hard and non-stationary, preventing the proposer from locking into fixed token sequences. Training Qwen3-4B and Qwen3-8B on mathematical reasoning via R-Zero, we find that vocabulary dropout sustains proposer diversity across lexical, semantic, and functional metrics throughout training, and yields solver improvements averaging +4.4 points at 8B, with the largest gains on competition-level benchmarks. Our findings suggest that explicit action-space constraints, analogous to the structural role that game rules play in classical self-play, can help sustain productive co-evolution in language. Vocabulary dropout is one simple instantiation of this principle.

Zijun Gao, Zhikun Xu, Xiao Ye et al.

Large language models (LLMs) often solve challenging math exercises yet fail to apply the concept right when the problem requires genuine understanding. Popular Reinforcement Learning with Verifiable Rewards (RLVR) pipelines reinforce final answers but provide little fine-grained conceptual signal, so models improve at pattern reuse rather than conceptual applications. We introduce CORE (Concept-Oriented REinforcement), an RL training framework that turns explicit concepts into a controllable supervision signal. Starting from a high-quality, low-contamination textbook resource that links verifiable exercises to concise concept descriptions, we run a sanity probe showing LLMs can restate definitions but fail concept-linked quizzes, quantifying the conceptual reasoning gap. CORE then (i) synthesizes concept-aligned quizzes, (ii) injects brief concept snippets during rollouts to elicit concept-primed trajectories, and (iii) reinforces conceptual reasoning via trajectory replacement after group failures, a lightweight forward-KL constraint that aligns unguided with concept-primed policies, or standard GRPO directly on concept-aligned quizzes. Across several models, CORE delivers consistent gains over vanilla and SFT baselines on both in-domain concept-exercise suites and diverse out-of-domain math benchmarks. CORE unifies direct training on concept-aligned quizzes and concept-injected rollouts under outcome regularization. It provides fine-grained conceptual supervision that bridges problem-solving competence and genuine conceptual reasoning, while remaining algorithm- and verifier-agnostic.

Aswin RRV, Jacob Dineen, Divij Handa et al.

The effectiveness of Reinforcement Learning (RL) in Large Language Models (LLMs) depends on the nature and diversity of the data used before and during RL. In particular, reasoning problems can often be approached in multiple ways that rely on different forms of reasoning, and exposure to only a limited range of such approaches in the training data may limit the effectiveness of RL. Motivated by this, we investigate using diverse self-generated data during mid-training as an intermediate step before RL training. Specifically, we adopt a bootstrapped data-generation framework guided by George Polya's problem-solving approaches for generating multiple variants of correct answers for each question in the training data, and then perform fine-tuning. We first provide a theoretical perspective on how mid-training on such data improves RL and explain how policy-gradient updates can incentivize combining multiple approaches. We then empirically demonstrate that RL-trained models initialized with our mid-training data achieve consistent improvements across various mathematical reasoning benchmarks and other OOD tasks like code generation and narrative reasoning. Overall, our investigative study shows that a language model learning multiple problem-solving approaches, through self-generated data helps subsequent RL.

Ben Zhou, Hongming Zhang, Sihao Chen et al.

Conceptual reasoning, the ability to reason in abstract and high-level perspectives, is key to generalization in human cognition. However, limited study has been done on large language models' capability to perform conceptual reasoning. In this work, we bridge this gap and propose a novel conceptualization framework that forces models to perform conceptual reasoning on abstract questions and generate solutions in a verifiable symbolic space. Using this framework as an analytical tool, we show that existing large language models fall short on conceptual reasoning, dropping 9% to 28% on various benchmarks compared to direct inference methods. We then discuss how models can improve since high-level abstract reasoning is key to unbiased and generalizable decision-making. We propose two techniques to add trustworthy induction signals by generating familiar questions with similar underlying reasoning paths and asking models to perform self-refinement. Experiments show that our proposed techniques improve models' conceptual reasoning performance by 8% to 11%, achieving a more robust reasoning system that relies less on inductive biases.

Yu Feng, Ben Zhou, Weidong Lin et al.

Predictive models often need to work with incomplete information in real-world tasks. Consequently, they must provide reliable probability or confidence estimation, especially in large-scale decision-making and planning tasks. Current large language models (LLMs) are insufficient for accurate estimations, but they can generate relevant factors that may affect the probabilities, produce coarse-grained probabilities when the information is more complete, and help determine which factors are relevant to specific downstream contexts. In this paper, we make use of these capabilities of LLMs to provide a significantly more accurate probabilistic estimation. We propose BIRD, a novel probabilistic inference framework that aligns a Bayesian network with LLM abductions and then estimates more accurate probabilities in a deduction step. We show BIRD provides reliable probability estimations that are 30% better than those provided directly by LLM baselines. These estimates further contribute to better and more trustworthy decision making.

Xiaodong Yu, Ben Zhou, Hao Cheng et al.

Existing math datasets evaluate the reasoning abilities of large language models (LLMs) by either using the final answer or the intermediate reasoning steps derived from static examples. However, the former approach fails to surface model's uses of shortcuts and wrong reasoning while the later poses challenges in accommodating alternative solutions. In this work, we seek to use symbolic programs as a means for automated evaluation if a model can consistently produce correct final answers across various inputs to the program. We begin by extracting programs for popular math datasets (GSM8K and MATH) using GPT4-o. For those executable programs verified using the original input-output pairs, they are found to encapsulate the proper reasoning required to solve the original text questions. We then prompt GPT4-o to generate new questions using alternative input-output pairs based the extracted program. We apply the resulting datasets to evaluate a collection of LLMs. In our experiments, we observe significant accuracy drops using our proposed evaluation compared with original static examples, suggesting the fragility of math reasoning in state-of-the-art LLMs.

Ruosen Li, Ziming Luo, Quan Zhang et al.

Large reasoning models (LRMs) achieve impressive reasoning capabilities by generating lengthy chain-of-thoughts, but this "overthinking" incurs high latency and cost without commensurate accuracy gains. In this work, we introduce AALC, a lightweight, accuracy-aware length reward integrated into reinforcement learning that dynamically balances correctness and brevity during training. By incorporating validation accuracy into the reward and employing a smooth, dynamically scheduled length penalty, AALC delays length penalty until target performance is met. Through extensive experiments across standard and out-of-distribution math benchmarks, we show that our approach reduces response length by over 50% while maintaining or even improving the original accuracy. Furthermore, qualitative analysis reveals that our method curbs redundant reasoning patterns such as excessive subgoal setting and verification, leading to structurally refined outputs rather than naive truncation. We also identify that efficiency gains are accompanied by reduced interpretability: models trained with AALC omit some narrative framing and explanatory context. These findings highlight the potential of reward-based strategies to guide LRMs toward more efficient, generalizable reasoning paths.

Jacob Dineen, Aswin RRV, Qin Liu et al.

Alignment of large language models (LLMs) with principles like helpfulness, honesty, and harmlessness typically relies on scalar rewards that obscure which objectives drive the training signal. We introduce QA-LIGN, which decomposes monolithic rewards into interpretable principle-specific evaluations through structured natural language programs. Models learn through a draft, critique, and revise pipeline, where symbolic evaluation against the rubrics provides transparent feedback for both initial and revised responses during GRPO training. Applied to uncensored Llama-3.1-8B-Instruct, QA-LIGN reduces attack success rates by up to 68.7% while maintaining a 0.67% false refusal rate, achieving Pareto optimal safety-helpfulness performance and outperforming both DPO and GRPO with state-of-the-art reward models given equivalent training. These results demonstrate that making reward signals interpretable and modular improves alignment effectiveness, suggesting transparency enhances LLM safety.

Longchao Da, Tiejin Chen, Zhuoheng Li et al.

The integration of generative artificial intelligence (GenAI) into transportation planning has the potential to revolutionize tasks such as demand forecasting, infrastructure design, policy evaluation, and traffic simulation. However, there is a critical need for a systematic framework to guide the adoption of GenAI in this interdisciplinary domain. In this survey, we, a multidisciplinary team of researchers spanning computer science and transportation engineering, present the first comprehensive framework for leveraging GenAI in transportation planning. Specifically, we introduce a new taxonomy that categorizes existing applications and methodologies into two perspectives: transportation planning tasks and computational techniques. From the transportation planning perspective, we examine the role of GenAI in automating descriptive, predictive, generative, simulation, and explainable tasks to enhance mobility systems. From the computational perspective, we detail advancements in data preparation, domain-specific fine-tuning, and inference strategies, such as retrieval-augmented generation and zero-shot learning tailored to transportation applications. Additionally, we address critical challenges, including data scarcity, explainability, bias mitigation, and the development of domain-specific evaluation frameworks that align with transportation goals like sustainability, equity, and system efficiency. This survey aims to bridge the gap between traditional transportation planning methodologies and modern AI techniques, fostering collaboration and innovation. By addressing these challenges and opportunities, we seek to inspire future research that ensures ethical, equitable, and impactful use of generative AI in transportation planning.

Ben Zhou, Sarthak Jain, Yi Zhang et al.

Large language models have been shown to suffer from reasoning inconsistency issues. That is, they fail more in situations unfamiliar to the training data, even though exact or very similar reasoning paths exist in more common cases that they can successfully solve. Such observations motivate us to propose methods that encourage models to understand the high-level and abstract reasoning processes during training instead of only the final answer. This way, models can transfer the exact solution to similar cases, regardless of their relevance to the pre-training data distribution. In this work, we propose SAL, a self-supervised analogical learning framework. SAL mimics the human analogy process and trains models to explicitly transfer high-quality symbolic solutions from cases that they know how to solve to other rare cases in which they tend to fail more. We show that the resulting models after SAL learning outperform base language models on a wide range of reasoning benchmarks, such as StrategyQA, GSM8K, and HotpotQA, by 2% to 20%. At the same time, we show that our model is more generalizable and controllable through analytical studies.

Xiao Ye, Shaswat Shrivastava, Zhaonan Li et al.

Large language models excel at many tasks but still struggle with consistent, robust reasoning. We introduce Cohort-based Consistency Learning (CC-Learn), a reinforcement learning framework that improves the reliability of LLM reasoning by training on cohorts of similar questions derived from shared programmatic abstractions. To enforce cohort-level consistency, we define a composite objective combining cohort accuracy, a retrieval bonus for effective problem decomposition, and a rejection penalty for trivial or invalid lookups that reinforcement learning can directly optimize, unlike supervised fine-tuning. Optimizing this reward guides the model to adopt uniform reasoning patterns across all cohort members. Experiments on challenging reasoning benchmarks (including ARC-Challenge and StrategyQA) show that CC-Learn boosts both accuracy and reasoning stability over pretrained and SFT baselines. These results demonstrate that cohort-level RL effectively enhances reasoning consistency in LLMs.

Zhikun Xu, Ming Shen, Jacob Dineen et al.

We introduce thoughts of words (ToW), a novel training-time data-augmentation method for next-word prediction. ToW views next-word prediction as a core reasoning task and injects fine-grained thoughts explaining what the next word should be and how it is related to the previous contexts in pre-training texts. Our formulation addresses two fundamental drawbacks of existing next-word prediction learning schemes: they induce factual hallucination and are inefficient for models to learn the implicit reasoning processes in raw texts. While there are many ways to acquire such thoughts of words, we explore the first step of acquiring ToW annotations through distilling from larger models. After continual pre-training with only 70K ToW annotations, we effectively improve models' reasoning performances by 7% to 9% on average and reduce model hallucination by up to 10%. At the same time, ToW is entirely agnostic to tasks and applications, introducing no additional biases on labels or semantics.

Adarsh Srinivasan, Jacob Dineen, Muhammad Umar Afzal et al.

Large language models in healthcare often miss critical emotional cues, delivering medically sound but emotionally flat advice. This is especially problematic in clinical contexts where patients are distressed and vulnerable, and require empathic communication to support safety, adherence, and trust. We present RECAP (Reflect-Extract-Calibrate-Align-Produce), an inference-time framework that adds structured emotional reasoning without retraining. By decomposing empathy into transparent appraisal-theoretic stages and exposing per-dimension Likert signals, RECAP produces nuanced, auditable responses. Across EmoBench, SECEU, and EQ-Bench, RECAP improves emotional reasoning by 22-28% on 8B models and 10-13% on larger models over zero-shot baselines. Clinician evaluations further confirm superior empathetic communication. RECAP shows that modular, theory-grounded prompting can systematically enhance emotional intelligence in medical AI while preserving the accountability required for deployment.

Ming Shen, Zhikun Xu, Jacob Dineen et al.

Large language models (LLMs) are typically pretrained with next-word prediction (NWP), which yields strong surface fluency but places limited pressure on models to form explicit reasoning before emitting tokens. We study whether shifting the supervision signal can better elicit explicit reasoning and, more broadly, strengthen models' general reasoning capability. We present BOttlenecked next-Word prediction (BOW), a RL formulation of NWP that inserts an intermediate reasoning bottleneck. Instead of predicting the next word directly from context, the policy model must first generate a next-word reasoning trajectory. A frozen scorer then assigns this trajectory a soft, distributional reward equal to the probability of the gold next token conditioned solely on the trajectory to guide the RL optimization. We also propose an optional L1-style regularizer on the reward to discourage "name-the-answer" shortcuts. Across ten benchmarks, a brief BOW adaptation phase on Qwen2.5-7B-Instruct and Llama3.1-8B-Instruct improves zero-shot reasoning and outperforms strong continual-pretraining baselines, including an RL variant with a hard, binary reward and a supervised finetuning approach with augmented data, by nearly 5% on average, while achieving the top result in 7 of 10 intrinsic NWP evaluations. These results indicate that BOW is a viable alternative to vanilla NWP, inducing explicit next-word reasoning and strengthening general reasoning ability.

Bangzheng Li, Fei Wang, Wenxuan Zhou et al. · microsoft-research

Vision-Language Models (VLMs) leverage aligned visual encoders to transform images into visual tokens, allowing them to be processed similarly to text by the backbone large language model (LLM). This unified input paradigm enables VLMs to excel in vision-language tasks such as visual question answering (VQA). To improve fine-grained visual reasoning, recent advancements in vision-language modeling introduce image cropping techniques that feed all encoded sub-images into the model. However, this approach significantly increases the number of visual tokens, leading to inefficiency and potential distractions for the LLM. To address the generalization challenges of image representation in VLMs, we propose a lightweight, universal framework that seamlessly integrates with existing VLMs to enhance their ability to process finegrained details. Our method leverages textual semantics to identify key visual areas, improving VQA performance without requiring any retraining of the VLM. Additionally, it incorporates textual signals into the visual encoding process, enhancing both efficiency and effectiveness. The proposed method, SEMCLIP, strengthens the visual understanding of a 7B VLM, LLaVA-1.5 by 3.3% on average across 7 benchmarks, and particularly by 5.3% on the challenging detailed understanding benchmark V*.

Zhikun Xu, Xiaodong Yu, Ben Zhou et al.

Recent large language models (LLMs) perform strongly on mathematical benchmarks yet often misapply lemmas, importing conclusions without validating assumptions. We formalize lemma$-$judging as a structured prediction task: given a statement and a candidate lemma, the model must output a precondition check and a conclusion$-$utility check, from which a usefulness decision is derived. We present RULES, which encodes this specification via a two$-$section output and trains with reinforcement learning plus section$-$aware loss masking to assign penalty to the section responsible for errors. Training and evaluation draw on diverse natural language and formal proof corpora; robustness is assessed with a held$-$out perturbation suite; and end$-$to$-$end evaluation spans competition$-$style, perturbation$-$aligned, and theorem$-$based problems across various LLMs. Results show consistent in$-$domain gains over both a vanilla model and a single$-$label RL baseline, larger improvements on applicability$-$breaking perturbations, and parity or modest gains on end$-$to$-$end tasks; ablations indicate that the two$-$section outputs and section$-$aware reinforcement are both necessary for robustness.

Xiao Ye, Jacob Dineen, Zhaonan Li et al.

Medical Large language models achieve strong scores on standard benchmarks; however, the transfer of those results to safe and reliable performance in clinical workflows remains a challenge. This survey reframes evaluation through a levels-of-autonomy lens (L0-L3), spanning informational tools, information transformation and aggregation, decision support, and supervised agents. We align existing benchmarks and metrics with the actions permitted at each level and their associated risks, making the evaluation targets explicit. This motivates a level-conditioned blueprint for selecting metrics, assembling evidence, and reporting claims, alongside directions that link evaluation to oversight. By centering autonomy, the survey moves the field beyond score-based claims toward credible, risk-aware evidence for real clinical use.

Qin Liu, Jacob Dineen, Yuxi Huang et al. · microsoft-research

Benchmarks are central to measuring the capabilities of large language models and guiding model development, yet widespread data leakage from pretraining corpora undermines their validity. Models can match memorized content rather than demonstrate true generalization, which inflates scores, distorts cross-model comparisons, and misrepresents progress. We introduce ArenaBencher, a model-agnostic framework for automatic benchmark evolution that updates test cases while preserving comparability. Given an existing benchmark and a diverse pool of models to be evaluated, ArenaBencher infers the core ability of each test case, generates candidate question-answer pairs that preserve the original objective, verifies correctness and intent with an LLM as a judge, and aggregates feedback from multiple models to select candidates that expose shared weaknesses. The process runs iteratively with in-context demonstrations that steer generation toward more challenging and diagnostic cases. We apply ArenaBencher to math problem solving, commonsense reasoning, and safety domains and show that it produces verified, diverse, and fair updates that uncover new failure modes, increase difficulty while preserving test objective alignment, and improve model separability. The framework provides a scalable path to continuously evolve benchmarks in step with the rapid progress of foundation models.

Bangzheng Li, Ben Zhou, Xingyu Fu et al.

Language models have shown impressive in-context-learning capabilities, which allow them to benefit from input prompts and perform better on downstream end tasks. Existing works investigate the mechanisms behind this observation, and propose label-agnostic prompt metrics that can better estimate end-task performances. One popular approach is using perplexity as a way to measure models' familiarity with the prompt. While showing consistent improvements on in-domain tasks, we found that familiarity metrics such as perplexity cannot accurately estimate performance in complicated situations such as task or domain transferring scenarios. In this work, we propose a revised measure called FamiCom, providing a more comprehensive measure for task-agnostic performance estimation. Specifically, FamiCom combines familiarity with \textit{complexity} -- the inherent difficulty of end tasks, which is an important factor missing from current metrics. Experiments show that FamiCom strongly correlates with end-task performances, producing a 0.85 Spearman's correlation, versus 0.43 of familiarity-only ones'. We further apply FamiCom to automatic prompt and demonstration selection, and outperform existing methods and baselines by more than 7.0% in accuracy.

Xingyu Fu, Ben Zhou, Sihao Chen et al.

Recent advances in multimodal large language models (LLMs) have shown extreme effectiveness in visual question answering (VQA). However, the design nature of these end-to-end models prevents them from being interpretable to humans, undermining trust and applicability in critical domains. While post-hoc rationales offer certain insight into understanding model behavior, these explanations are not guaranteed to be faithful to the model. In this paper, we address these shortcomings by introducing an interpretable by design model that factors model decisions into intermediate human-legible explanations, and allows people to easily understand why a model fails or succeeds. We propose the Dynamic Clue Bottleneck Model ( (DCLUB), a method that is designed towards an inherently interpretable VQA system. DCLUB provides an explainable intermediate space before the VQA decision and is faithful from the beginning, while maintaining comparable performance to black-box systems. Given a question, DCLUB first returns a set of visual clues: natural language statements of visually salient evidence from the image, and then generates the output based solely on the visual clues. To supervise and evaluate the generation of VQA explanations within DCLUB, we collect a dataset of 1.7k reasoning-focused questions with visual clues. Evaluations show that our inherently interpretable system can improve 4.64% over a comparable black-box system in reasoning-focused questions while preserving 99.43% of performance on VQA-v2.

Ben Zhou, Kyle Richardson, Qiang Ning et al.

We propose TRACIE, a novel temporal reasoning dataset that evaluates the degree to which systems understand implicit events -- events that are not mentioned explicitly in natural language text but can be inferred from it. This introduces a new challenge in temporal reasoning research, where prior work has focused on explicitly mentioned events. Human readers can infer implicit events via commonsense reasoning, resulting in a more comprehensive understanding of the situation and, consequently, better reasoning about time. We find, however, that state-of-the-art models struggle when predicting temporal relationships between implicit and explicit events. To address this, we propose a neuro-symbolic temporal reasoning model, SYMTIME, which exploits distant supervision signals from large-scale text and uses temporal rules to combine start times and durations to infer end times. SYMTIME outperforms strong baseline systems on TRACIE by 5%, and by 11% in a zero prior knowledge training setting. Our approach also generalizes to other temporal reasoning tasks, as evidenced by a gain of 1%-9% on MATRES, an explicit event benchmark.

Ben Zhou, Qiang Ning, Daniel Khashabi et al.

Temporal common sense (e.g., duration and frequency of events) is crucial for understanding natural language. However, its acquisition is challenging, partly because such information is often not expressed explicitly in text, and human annotation on such concepts is costly. This work proposes a novel sequence modeling approach that exploits explicit and implicit mentions of temporal common sense, extracted from a large corpus, to build TACOLM, a temporal common sense language model. Our method is shown to give quality predictions of various dimensions of temporal common sense (on UDST and a newly collected dataset from RealNews). It also produces representations of events for relevant tasks such as duration comparison, parent-child relations, event coreference and temporal QA (on TimeBank, HiEVE and MCTACO) that are better than using the standard BERT. Thus, it will be an important component of temporal NLP.

Muhao Chen, Weijia Shi, Ben Zhou et al.

Much research effort has been put to multilingual knowledge graph (KG) embedding methods to address the entity alignment task, which seeks to match entities in different languagespecific KGs that refer to the same real-world object. Such methods are often hindered by the insufficiency of seed alignment provided between KGs. Therefore, we propose an incidentally supervised model, JEANS , which jointly represents multilingual KGs and text corpora in a shared embedding scheme, and seeks to improve entity alignment with incidental supervision signals from text. JEANS first deploys an entity grounding process to combine each KG with the monolingual text corpus. Then, two learning processes are conducted: (i) an embedding learning process to encode the KG and text of each language in one embedding space, and (ii) a selflearning based alignment learning process to iteratively induce the matching of entities and that of lexemes between embeddings. Experiments on benchmark datasets show that JEANS leads to promising improvement on entity alignment with incidental supervision, and significantly outperforms state-of-the-art methods that solely rely on internal information of KGs.

Matt Gardner, Yoav Artzi, Victoria Basmova et al.

Standard test sets for supervised learning evaluate in-distribution generalization. Unfortunately, when a dataset has systematic gaps (e.g., annotation artifacts), these evaluations are misleading: a model can learn simple decision rules that perform well on the test set but do not capture a dataset's intended capabilities. We propose a new annotation paradigm for NLP that helps to close systematic gaps in the test data. In particular, after a dataset is constructed, we recommend that the dataset authors manually perturb the test instances in small but meaningful ways that (typically) change the gold label, creating contrast sets. Contrast sets provide a local view of a model's decision boundary, which can be used to more accurately evaluate a model's true linguistic capabilities. We demonstrate the efficacy of contrast sets by creating them for 10 diverse NLP datasets (e.g., DROP reading comprehension, UD parsing, IMDb sentiment analysis). Although our contrast sets are not explicitly adversarial, model performance is significantly lower on them than on the original test sets---up to 25\% in some cases. We release our contrast sets as new evaluation benchmarks and encourage future dataset construction efforts to follow similar annotation processes.

Ben Zhou, Daniel Khashabi, Qiang Ning et al.

Understanding time is crucial for understanding events expressed in natural language. Because people rarely say the obvious, it is often necessary to have commonsense knowledge about various temporal aspects of events, such as duration, frequency, and temporal order. However, this important problem has so far received limited attention. This paper systematically studies this temporal commonsense problem. Specifically, we define five classes of temporal commonsense, and use crowdsourcing to develop a new dataset, MCTACO, that serves as a test set for this task. We find that the best current methods used on MCTACO are still far behind human performance, by about 20%, and discuss several directions for improvement. We hope that the new dataset and our study here can foster more future research on this topic.

Ben Zhou, Daniel Khashabi, Chen-Tse Tsai et al.

The problem of entity-typing has been studied predominantly in supervised learning fashion, mostly with task-specific annotations (for coarse types) and sometimes with distant supervision (for fine types). While such approaches have strong performance within datasets, they often lack the flexibility to transfer across text genres and to generalize to new type taxonomies. In this work we propose a zero-shot entity typing approach that requires no annotated data and can flexibly identify newly defined types. Given a type taxonomy defined as Boolean functions of FREEBASE "types", we ground a given mention to a set of type-compatible Wikipedia entries and then infer the target mention's types using an inference algorithm that makes use of the types of these entries. We evaluate our system on a broad range of datasets, including standard fine-grained and coarse-grained entity typing datasets, and also a dataset in the biological domain. Our system is shown to be competitive with state-of-the-art supervised NER systems and outperforms them on out-of-domain datasets. We also show that our system significantly outperforms other zero-shot fine typing systems.

Qiang Ning, Ben Zhou, Zhili Feng et al.

Automatic extraction of temporal information in text is an important component of natural language understanding. It involves two basic tasks: (1) Understanding time expressions that are mentioned explicitly in text (e.g., February 27, 1998 or tomorrow), and (2) Understanding temporal information that is conveyed implicitly via relations. In this paper, we introduce CogCompTime, a system that has these two important functionalities. It incorporates the most recent progress, achieves state-of-the-art performance, and is publicly available.1 We believe that this demo will be useful for multiple time-aware applications and provide valuable insight for future research in temporal understanding.