COMPASS: A Multi-Turn Benchmark for Tool-Mediated Planning & Preference Optimization

Real-world large language model (LLM) agents must master strategic tool use and user preference optimization through multi-turn interactions to assist users with complex planning tasks. We introduce COMPASS (Constrained Optimization through Multi-turn Planning and Strategic Solutions), a benchmark that evaluates agents on realistic travel-planning scenarios. We cast travel planning as a constrained preference optimization problem, where agents must satisfy hard constraints while simultaneously optimizing soft user preferences. To support this, we build a realistic travel database covering transportation, accommodation, and ticketing for 20 U.S. National Parks, along with a comprehensive tool ecosystem that mirrors commercial booking platforms. Evaluating state-of-the-art models, we uncover two critical gaps: (i) an acceptable-optimal gap, where agents reliably meet constraints but fail to optimize preferences, and (ii) a plan-coordination gap, where performance collapses on multi-service (flight and hotel) coordination tasks, especially for open-source models. By grounding reasoning and planning in a practical, user-facing domain, COMPASS provides a benchmark that directly measures an agent's ability to optimize user preferences in realistic tasks, bridging theoretical advances with real-world impact.

Checklists Are Better Than Reward Models For Aligning Language Models

Language models must be adapted to understand and follow user instructions. Reinforcement learning is widely used to facilitate this -- typically using fixed criteria such as "helpfulness" and "harmfulness". In our work, we instead propose using flexible, instruction-specific criteria as a means of broadening the impact that reinforcement learning can have in eliciting instruction following. We propose "Reinforcement Learning from Checklist Feedback" (RLCF). From instructions, we extract checklists and evaluate how well responses satisfy each item - using both AI judges and specialized verifier programs - then combine these scores to compute rewards for RL. We compare RLCF with other alignment methods applied to a strong instruction following model (Qwen2.5-7B-Instruct) on five widely-studied benchmarks -- RLCF is the only method to improve performance on every benchmark, including a 4-point boost in hard satisfaction rate on FollowBench, a 6-point increase on InFoBench, and a 3-point rise in win rate on Arena-Hard. These results establish checklist feedback as a key tool for improving language models' support of queries that express a multitude of needs.

C2-Evo: Co-Evolving Multimodal Data and Model for Self-Improving Reasoning

Recent advances in multimodal large language models (MLLMs) have shown impressive reasoning capabilities. However, further enhancing existing MLLMs necessitates high-quality vision-language datasets with carefully curated task complexities, which are both costly and challenging to scale. Although recent self-improving models that iteratively refine themselves offer a feasible solution, they still suffer from two core challenges: (i) most existing methods augment visual or textual data separately, resulting in discrepancies in data complexity (e.g., over-simplified diagrams paired with redundant textual descriptions); and (ii) the evolution of data and models is also separated, leading to scenarios where models are exposed to tasks with mismatched difficulty levels. To address these issues, we propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities. Specifically, given a base dataset and a base model, C2-Evo enhances them by a cross-modal data evolution loop and a data-model evolution loop. The former loop expands the base dataset by generating complex multimodal problems that combine structured textual sub-problems with iteratively specified geometric diagrams, while the latter loop adaptively selects the generated problems based on the performance of the base model, to conduct supervised fine-tuning and reinforcement learning alternately. Consequently, our method continuously refines its model and training data, and consistently obtains considerable performance gains across multiple mathematical reasoning benchmarks. Our code, models, and datasets will be released.

PhyBlock: A Progressive Benchmark for Physical Understanding and Planning via 3D Block Assembly

While vision-language models (VLMs) have demonstrated promising capabilities in reasoning and planning for embodied agents, their ability to comprehend physical phenomena, particularly within structured 3D environments, remains severely limited. To close this gap, we introduce PhyBlock, a progressive benchmark designed to assess VLMs on physical understanding and planning through robotic 3D block assembly tasks. PhyBlock integrates a novel four-level cognitive hierarchy assembly task alongside targeted Visual Question Answering (VQA) samples, collectively aimed at evaluating progressive spatial reasoning and fundamental physical comprehension, including object properties, spatial relationships, and holistic scene understanding. PhyBlock includes 2600 block tasks (400 assembly tasks, 2200 VQA tasks) and evaluates models across three key dimensions: partial completion, failure diagnosis, and planning robustness. We benchmark 21 state-of-the-art VLMs, highlighting their strengths and limitations in physically grounded, multi-step planning. Our empirical findings indicate that the performance of VLMs exhibits pronounced limitations in high-level planning and reasoning capabilities, leading to a notable decline in performance for the growing complexity of the tasks. Error analysis reveals persistent difficulties in spatial orientation and dependency reasoning. Surprisingly, chain-of-thought prompting offers minimal improvements, suggesting spatial tasks heavily rely on intuitive model comprehension. We position PhyBlock as a unified testbed to advance embodied reasoning, bridging vision-language understanding and real-world physical problem-solving.

ScaleLong: A Multi-Timescale Benchmark for Long Video Understanding

Although long-video understanding demands that models capture hierarchical temporal information -- from clip (seconds) and shot (tens of seconds) to event (minutes) and story (hours) -- existing benchmarks either neglect this multi-scale design or scatter scale-specific questions across different videos, preventing direct comparison of model performance across timescales on the same content. To address this, we introduce ScaleLong, the first benchmark to disentangle these factors by embedding questions targeting four hierarchical timescales -- clip (seconds), shot (tens of seconds), event (minutes), and story (hours) -- all within the same video content. This within-content multi-timescale questioning design enables direct comparison of model performance across timescales on identical videos. ScaleLong features 269 long videos (avg.\ 86\,min) from 5 main categories and 36 sub-categories, with 4--8 carefully designed questions, including at least one question for each timescale. Evaluating 23 MLLMs reveals a U-shaped performance curve, with higher accuracy at the shortest and longest timescales and a dip at intermediate levels. Furthermore, ablation studies show that increased visual token capacity consistently enhances reasoning across all timescales. ScaleLong offers a fine-grained, multi-timescale benchmark for advancing MLLM capabilities in long-video understanding. The code and dataset are available https://github.com/multimodal-art-projection/ScaleLong.

Can LLMs Reason Abstractly Over Math Word Problems Without CoT? Disentangling Abstract Formulation From Arithmetic Computation

Final-answer-based metrics are commonly used for evaluating large language models (LLMs) on math word problems, often taken as proxies for reasoning ability. However, such metrics conflate two distinct sub-skills: abstract formulation (capturing mathematical relationships using expressions) and arithmetic computation (executing the calculations). Through a disentangled evaluation on GSM8K and SVAMP, we find that the final-answer accuracy of Llama-3 and Qwen2.5 (1B-32B) without CoT is overwhelmingly bottlenecked by the arithmetic computation step and not by the abstract formulation step. Contrary to the common belief, we show that CoT primarily aids in computation, with limited impact on abstract formulation. Mechanistically, we show that these two skills are composed conjunctively even in a single forward pass without any reasoning steps via an abstract-then-compute mechanism: models first capture problem abstractions, then handle computation. Causal patching confirms these abstractions are present, transferable, composable, and precede computation. These behavioural and mechanistic findings highlight the need for disentangled evaluation to accurately assess LLM reasoning and to guide future improvements.

Stochastic Chameleons: Irrelevant Context Hallucinations Reveal Class-Based (Mis)Generalization in LLMs

The widespread success of large language models (LLMs) on NLP benchmarks has been accompanied by concerns that LLMs function primarily as stochastic parrots that reproduce texts similar to what they saw during pre-training, often erroneously. But what is the nature of their errors, and do these errors exhibit any regularities? In this work, we examine irrelevant context hallucinations, in which models integrate misleading contextual cues into their predictions. Through behavioral analysis, we show that these errors result from a structured yet flawed mechanism that we term class-based (mis)generalization, in which models combine abstract class cues with features extracted from the query or context to derive answers. Furthermore, mechanistic interpretability experiments on Llama-3, Mistral, and Pythia across 39 factual recall relation types reveal that this behavior is reflected in the model's internal computations: (i) abstract class representations are constructed in lower layers before being refined into specific answers in higher layers, (ii) feature selection is governed by two competing circuits -- one prioritizing direct query-based reasoning, the other incorporating contextual cues -- whose relative influences determine the final output. Our findings provide a more nuanced perspective on the stochastic parrot argument: through form-based training, LLMs can exhibit generalization leveraging abstractions, albeit in unreliable ways based on contextual cues -- what we term stochastic chameleons.

SCAR: Shapley Credit Assignment for More Efficient RLHF

Reinforcement Learning from Human Feedback (RLHF) is a widely used technique for aligning Large Language Models (LLMs) with human preferences, yet it often suffers from sparse reward signals, making effective credit assignment challenging. In typical setups, the reward model provides a single scalar score for an entire generated sequence, offering little insight into which token or span-level decisions were responsible for the outcome. To address this, we propose Shapley Credit Assignment Rewards (SCAR), a novel method that leverages Shapley values in cooperative game theory. SCAR distributes the total sequence-level reward among constituent tokens or text spans based on their principled marginal contributions. This creates dense reward signals, crucially, without necessitating the training of auxiliary critique models or recourse to fine-grained human annotations at intermediate generation stages. Unlike prior dense reward methods, SCAR offers a game-theoretic foundation for fair credit attribution. Theoretically, we demonstrate that SCAR preserves the original optimal policy, and empirically, across diverse tasks including sentiment control, text summarization, and instruction tuning, we show that SCAR converges significantly faster and achieves higher final reward scores compared to standard RLHF and attention-based dense reward baselines. Our findings suggest that SCAR provides a more effective and theoretically sound method for credit assignment in RLHF, leading to more efficient alignment of LLMs.

Ground-R1: Incentivizing Grounded Visual Reasoning via Reinforcement Learning

Large Vision-Language Models (LVLMs) have demonstrated impressive general capabilities across a wide range of multi-modal tasks. However, the reasoning processes of LVLMs often suffer from unreliable outputs and limited interpretability. To address this, grounded visual reasoning has emerged as a promising paradigm that enforces responses anchored on salient visual evidence regions. However, existing approaches typically rely on costly supervision such as bounding box annotations, chain-of-thought rationale or external tool calls, limiting their scalability. In this work, we propose Ground-R1, a reinforcement learning framework that enables grounded visual reasoning without requiring explicit evidence or rationale annotations. Ground-R1 consists of a grounding phase that generates evidence region rollouts based on format constraints, and an answering phase that produces responses guided by both answer correctness and format adherence rewards. Extensive experiments across multiple visual reasoning benchmarks manifest that Ground-R1 achieves superior performance and exhibits emergent cognitive behaviors such as uncertainty awareness, spatial perception, and iterative refinement, offering a scalable and interpretable alternative to existing approaches.

Breaking Down Video LLM Benchmarks: Knowledge, Spatial Perception, or True Temporal Understanding?

Existing video understanding benchmarks often conflate knowledge-based and purely image-based questions, rather than clearly isolating a model's temporal reasoning ability, which is the key aspect that distinguishes video understanding from other modalities. We identify two major limitations that obscure whether higher scores truly indicate stronger understanding of the dynamic content in videos: (1) strong language priors, where models can answer questions without watching the video; and (2) shuffling invariance, where models maintain similar performance on certain questions even when video frames are temporally shuffled. To alleviate these issues, we propose VBenchComp, an automated pipeline that categorizes questions into different domains: LLM-Answerable, Semantic, and Temporal. Specifically, LLM-Answerable questions can be answered without viewing the video; Semantic questions remain answerable even when the video frames are shuffled; and Temporal questions require understanding the correct temporal order of frames. The rest of the questions are labeled as Others. This can enable fine-grained evaluation of different capabilities of a video LLM. Our analysis reveals nuanced model weaknesses that are hidden by traditional overall scores, and we offer insights and recommendations for designing future benchmarks that more accurately assess video LLMs.