Recent research has explored the constrained generation capabilities of Large Language Models (LLMs) when explicitly prompted by few task-specific requirements. In contrast, we introduce Large-Scale Constraint Generation (LSCG), a new problem that evaluates whether LLMs can parse a large, fine-grained, generic list of constraints. To examine the LLMs' ability to handle an increasing number constraints, we create a practical instance of LSCG, called Words Checker. In Words Checker, we evaluate the impact of model characteristics (e.g., size, family) and steering techniques (e.g., Simple Prompt, Chain of Thought, Best of N) on performance. We also propose FoCusNet, a small and dedicated model that parses the original list of constraints into a smaller subset, helping the LLM focus on relevant constraints. Experiments reveal that existing solutions suffer a significant performance drop as the number of constraints increases, with FoCusNet showing an 8-13% accuracy boost.

Context-grounded hallucinations are cases where model outputs contain information not verifiable against the source text. We study the applicability of LLMs for localizing such hallucinations, as a more practical alternative to existing complex evaluation pipelines. In the absence of established benchmarks for meta-evaluation of hallucinations localization, we construct one tailored to LLMs, involving a challenging human annotation of over 1,000 examples. We complement the benchmark with an LLM-based evaluation protocol, verifying its quality in a human evaluation. Since existing representations of hallucinations limit the types of errors that can be expressed, we propose a new representation based on free-form textual descriptions, capturing the full range of possible errors. We conduct a comprehensive study, evaluating four large-scale LLMs, which highlights the benchmark's difficulty, as the best model achieves an F1 score of only 0.67. Through careful analysis, we offer insights into optimal prompting strategies for the task and identify the main factors that make it challenging for LLMs: (1) a tendency to incorrectly flag missing details as inconsistent, despite being instructed to check only facts in the output; and (2) difficulty with outputs containing factually correct information absent from the source - and thus not verifiable - due to alignment with the model's parametric knowledge.

Chain-of-Thought (CoT) prompting helps Large Language Models (LLMs) tackle complex reasoning by eliciting explicit step-by-step rationales. However, CoT's verbosity increases latency and memory usage and may propagate early errors across long chains. We propose the Reasoning Capsule (R-Capsule), a framework that aims to combine the efficiency of latent reasoning with the transparency of explicit CoT. The core idea is to compress the high-level plan into a small set of learned latent tokens (a Reasoning Capsule) while keeping execution steps lightweight or explicit. This hybrid approach is inspired by the Information Bottleneck (IB) principle, where we encourage the capsule to be approximately minimal yet sufficient for the task. Minimality is encouraged via a low-capacity bottleneck, which helps improve efficiency. Sufficiency is encouraged via a dual objective: a primary task loss for answer accuracy and an auxiliary plan-reconstruction loss that encourages the capsule to faithfully represent the original textual plan. The reconstruction objective helps ground the latent space, thereby improving interpretability and reducing the use of uninformative shortcuts. Our framework strikes a balance between efficiency, accuracy, and interpretability, thereby reducing the visible token footprint of reasoning while maintaining or improving accuracy on complex benchmarks. Our codes are available at: https://anonymous.4open.science/r/Reasoning-Capsule-7BE0

The growing demand for large language models (LLMs) with tunable reasoning capabilities in many real-world applications highlights a critical need for methods that can efficiently produce a spectrum of models balancing reasoning depth and computational cost. Model merging has emerged as a promising, training-free technique to address this challenge by arithmetically combining the weights of a general-purpose model with a specialized reasoning model. While various merging techniques exist, their potential to create a spectrum of models with fine-grained control over reasoning abilities remains largely unexplored. This work presents a large-scale empirical study evaluating a range of model merging techniques across multiple reasoning benchmarks. We systematically vary merging strengths to construct accuracy-efficiency curves, providing the first comprehensive view of the tunable performance landscape. Our findings reveal that model merging offers an effective and controllable method for calibrating the trade-off between reasoning accuracy and token efficiency, even when parent models have highly divergent weight spaces. Crucially, we identify instances of Pareto Improvement, where a merged model achieves both higher accuracy and lower token consumption than one of its parents. Our study provides the first comprehensive analysis of this tunable space, offering practical guidelines for creating LLMs with specific reasoning profiles to meet diverse application demands. Large language models (LLMs) have become powerful, general-purpose tools, capable of addressing a wide array of tasks through a unified token generation process (Brown et al., 2020; Achiam et al., 2023; Liu et al., 2024).

Diffusion Transformers (DiT)-based video generation models with 3D full attention exhibit strong generative capabilities. Trajectory control represents a user-friendly task in the field of controllable video generation. However, existing methods either require substantial training resources or are specifically designed for U-Net, do not take advantage of the superior performance of DiT. To address these issues, we propose DiTraj, a simple but effective training-free framework for trajectory control in text-to-video generation, tailored for DiT. Specifically, first, to inject the object's trajectory, we propose foreground-background separation guidance: we use the Large Language Model (LLM) to convert user-provided prompts into foreground and background prompts, which respectively guide the generation of foreground and background regions in the video. Then, we analyze 3D full attention and explore the tight correlation between inter-token attention scores and position embedding. Based on this, we propose inter-frame Spatial-Temporal Decoupled 3D-RoPE (STD-RoPE). By modifying only foreground tokens' position embedding, STD-RoPE eliminates their cross-frame spatial discrepancies, strengthening cross-frame attention among them and thus enhancing trajectory control. Additionally, we achieve 3D-aware trajectory control by regulating the density of position embedding. Extensive experiments demonstrate that our method outperforms previous methods in both video quality and trajectory controllability.

Recent neural solvers have demonstrated promising performance in learning to solve routing problems. However, existing studies are primarily based on one-off training on one or a set of predefined problem distributions and scales, i.e., tasks. When a new task arises, they typically rely on either zero-shot generalization, which may be poor due to the discrepancies between the new task and the training task(s), or fine-tuning the pretrained solver on the new task, which possibly leads to catastrophic forgetting of knowledge acquired from previous tasks. This paper explores a novel lifelong learning paradigm for neural VRP solvers, where multiple tasks with diverse distributions and scales arise sequentially over time. Solvers are required to effectively and efficiently learn to solve new tasks while maintaining their performance on previously learned tasks. Consequently, a novel framework called Lifelong Learning Router with Behavior Consolidation (LLR-BC) is proposed. LLR-BC consolidates prior knowledge effectively by aligning behaviors of the solver trained on a new task with the buffered ones in a decision-seeking way. To encourage more focus on crucial experiences, LLR-BC assigns greater consolidated weights to decisions with lower confidence. Extensive experiments on capacitated vehicle routing problems and traveling salesman problems demonstrate LLR-BC's effectiveness in training high-performance neural solvers in a lifelong learning setting, addressing the catastrophic forgetting issue, maintaining their plasticity, and improving zero-shot generalization ability.

Achieving generalizable bimanual manipulation requires systems that can learn efficiently from minimal human input while adapting to real-world uncertainties and diverse embodiments. Existing approaches face a dilemma: imitation policy learning demands extensive demonstrations to cover task variations, while modular methods often lack flexibility in dynamic scenes. We introduce VLBiMan, a framework that derives reusable skills from a single human example through task-aware decomposition, preserving invariant primitives as anchors while dynamically adapting adjustable components via vision-language grounding. This adaptation mechanism resolves scene ambiguities caused by background changes, object repositioning, or visual clutter without policy retraining, leveraging semantic parsing and geometric feasibility constraints. Moreover, the system inherits human-like hybrid control capabilities, enabling mixed synchronous and asynchronous use of both arms. Extensive experiments validate VLBiMan across tool-use and multi-object tasks, demonstrating: (1) a drastic reduction in demonstration requirements compared to imitation baselines, (2) compositional generalization through atomic skill splicing for long-horizon tasks, (3) robustness to novel but semantically similar objects and external disturbances, and (4) strong cross-embodiment transfer, showing that skills learned from human demonstrations can be instantiated on different robotic platforms without retraining. By bridging human priors with vision-language anchored adaptation, our work takes a step toward practical and versatile dual-arm manipulation in unstructured settings.

Chain-of-Thought (CoT) prompting often improves classification accuracy, but it introduces a significant throughput penalty with rationale generation (Wei et al., 2022; Cheng and Van Durme, 2024). To resolve this trade-off, we introduce Dual-Head Reasoning Distillation (DHRD), a simple training method for decoder-only language models (LMs) that adds (i) a pooled classification head used during training and inference and (ii) a reasoning head supervised by teacher rationales used only in training. We train with a loss function that is a weighted sum of label cross-entropy and token-level LM loss over input-plus-rationale sequences. On seven SuperGLUE tasks, DHRD yields relative gains of 0.65-5.47% over pooled baselines, with notably larger gains on entailment/causal tasks. Since we disable the reasoning head at test time, inference throughput matches pooled classifiers and exceeds CoT decoding on the same backbones by 96-142 times in QPS.

While diffusion language models (DLMs) have achieved competitive performance in text generation, improving their reasoning ability with reinforcement learning remains an active research area. Here, we introduce d2, a reasoning framework tailored for masked DLMs. Central to our framework is a new policy gradient algorithm that relies on properties of masking to accurately estimate the likelihoods of sampling trajectories. Our estimators trade off computation for approximation accuracy in an analytically tractable manner, and are particularly effective for DLMs that support any-order likelihood estimation. We characterize and study this property in popular DLMs and show that it is key for efficient diffusion-based reasoning. Empirically, d2 significantly improves over previous diffusion reasoning frameworks using only RL (without relying on supervised fine-tuning), and sets a new state-of-the-art performance for DLMs on logical reasoning tasks (Countdown and Sudoku) and math reasoning benchmarks (GSM8K and MATH500).

Purpose: To develop a retrieval-augmented generation (RAG) system powered by LLaMA-4 109B for automated, protocol-aware, and interpretable evaluation of radiotherapy treatment plans. Methods and Materials: We curated a multi-protocol dataset of 614 radiotherapy plans across four disease sites and constructed a knowledge base containing normalized dose metrics and protocol-defined constraints. The RAG system integrates three core modules: a retrieval engine optimized across five SentenceTransformer backbones, a percentile prediction component based on cohort similarity, and a clinical constraint checker. These tools are directed by a large language model (LLM) using a multi-step prompt-driven reasoning pipeline to produce concise, grounded evaluations. Results: Retrieval hyperparameters were optimized using Gaussian Process on a scalarized loss function combining root mean squared error (RMSE), mean absolute error (MAE), and clinically motivated accuracy thresholds. The best configuration, based on all-MiniLM-L6-v2, achieved perfect nearest-neighbor accuracy within a 5-percentile-point margin and a sub-2pt MAE. When tested end-to-end, the RAG system achieved 100% agreement with the computed values by standalone retrieval and constraint-checking modules on both percentile estimates and constraint identification, confirming reliable execution of all retrieval, prediction and checking steps. Conclusion: Our findings highlight the feasibility of combining structured population-based scoring with modular tool-augmented reasoning for transparent, scalable plan evaluation in radiation therapy. The system offers traceable outputs, minimizes hallucination, and demonstrates robustness across protocols. Future directions include clinician-led validation, and improved domain-adapted retrieval models to enhance real-world integration.