We train humanoid policies that compliantly respond to external forces while tracking a reference motion. The desired force-displacement relationship is modulated by a'stiffness' input at deployment time, and a single policy learns to realize a wide range of stiffnesses. In the images, the reference motion is visualized in blue, and the approximate external force on the robot is illustrated by the red arrows. Abstract-- We introduce SoftMimic, a framework for learning compliant whole-body control policies for humanoid robots from example motions. Imitating human motions with reinforcement learning allows humanoids to quickly learn new skills, but existing methods incentivize stiff control that aggressively corrects deviations from a reference motion, leading to brittle and unsafe behavior when the robot encounters unexpected contacts. In contrast, SoftMimic enables robots to respond compliantly to external forces while maintaining balance and posture. Our approach leverages an inverse kinematics solver to generate an augmented dataset of feasible compliant motions, which we use to train a reinforcement learning policy. By rewarding the policy for matching compliant responses rather than rigidly tracking the reference motion, SoftMimic learns to absorb disturbances and generalize to varied tasks from a single motion clip. I. INTRODUCTION A major goal in humanoid robotics is to build agents capable of performing a vast range of tasks humans execute in everyday environments. A promising avenue towards this goal is to leverage large-scale human motion capture data, enabling robots to learn human-like behaviors through imitation [1]. All authors are with the Improbable AI Lab, Massachusetts Institute of Technology, USA.

Scaled post-training now drives many of the largest capability gains in language models (LMs), yet its effect on pretrained knowledge remains poorly understood. Not all forgetting is equal: Forgetting one fact (e.g., a U.S. president or an API call) does not "average out" by recalling another. Hence, we propose a sample-wise paradigm to measure what is forgotten and when backward transfer occurs. Our metric counts 1->0 transitions (correct before post-training, incorrect after) to quantify forgetting and 0->1 transitions to quantify backward transfer. Traditional task averages conflate these effects and obscure large changes. For multiple-choice benchmarks, we add chance-adjusted variants that subtract the expected contribution of random guessing from pre- and post-training accuracies. We apply this framework across post-training stages, model sizes, and data scales. Our large-scale analysis shows that: (1) Domain-continual pretraining induces moderate forgetting with low-to-moderate backward transfer; (2) RL/SFT post-training applied to base models and Instruction tuning yields moderate-to-large backward transfer on math and logic with overall low-to-moderate forgetting; (3) Applying RL/SFT to instruction-tuned models is sensitive on data scale: at small scales, both forgetting and backward transfer are small; at larger scales, effects are mixed and warrant further study with better controls; (4) Model merging does not reliably mitigate forgetting. Overall, our framework offers a practical yardstick for mapping how post-training alters pretrained knowledge at scale -- enabling progress towards generally capable AI systems.

Effective code generation with language models hinges on two critical factors: accurately understanding the intent of the prompt and generating code that applies algorithmic reasoning to produce correct solutions capable of passing diverse test cases while adhering to the syntax of the target programming language. Unlike other language tasks, code generation requires more than accurate token prediction; it demands comprehension of solution-level and structural relationships rather than merely generating the most likely tokens. very large language model (VLLM) are capable of generating detailed steps toward the correct solution of complex tasks where reasoning is crucial in solving the problem. Such reasoning capabilities may be absent in smaller language models. Therefore, in this work, we distill the reasoning capabilities of a VLLM into a smaller, more efficient model that is faster and cheaper to deploy. Our approach trains the model to emulate the reasoning and problem-solving abilities of the VLLM by learning to identify correct solution pathways and establishing a structural correspondence between problem definitions and potential solutions through a novel method of structure-aware loss optimization. This enables the model to transcend token-level generation and to deeply grasp the overarching structure of solutions for given problems. Experimental results show that our fine-tuned model, developed through a cheap and simple to implement process, significantly outperforms our baseline model in terms of pass@1, average data flow, and average syntax match metrics across the MBPP, MBPP Plus, and HumanEval benchmarks.

We propose an extension of the decoder Transformer that conditions its generative process on random latent variables which are learned without supervision thanks to a variational procedure. Experimental evaluations show that allowing such a conditioning translates into substantial improvements on downstream tasks.

Traditional knowledge graphs are constrained by fixed ontologies that organize concepts within rigid hierarchical structures. The root cause lies in treating domains as implicit context rather than as explicit, reasoning-level components. To overcome these limitations, we propose the Domain-Contextualized Concept Graph (CDC), a novel knowledge modeling framework that elevates domains to first-class elements of conceptual representation. CDC adopts a C-D-C triple structure - - where domain specifications serve as dynamic classification dimensions defined on demand. Grounded in a cognitive-linguistic isomorphic mapping principle, CDC operationalizes how humans understand concepts through contextual frames. We formalize more than twenty standardized relation predicates (structural, logical, cross-domain, and temporal) and implement CDC in Prolog for full inference capability. Case studies in education, enterprise knowledge systems, and technical documentation demonstrate that CDC enables context-aware reasoning, cross-domain analogy, and personalized knowledge modeling - capabilities unattainable under traditional ontology-based frameworks.

Abstract--Real-world multi-agent systems may require ad hoc teaming, where an agent must coordinate with other previously unseen teammates to solve a task in a zero-shot manner . Prior work often either selects a pretrained policy based on an inferred model of the new teammates or pretrains a single policy that is robust to potential teammates. Instead, we propose to leverage all pretrained policies in a zero-shot transfer setting. We formalize this problem as an ad hoc multi-agent Markov decision process and present a solution that uses two key ideas, generalized policy improvement and difference rewards, for efficient and effective knowledge transfer between different teams. We empirically demonstrate that our algorithm, Generalized Policy improvement for Ad hoc T eaming (GPA T), successfully enables zero-shot transfer to new teams in three simulated environments: cooperative foraging, predator-prey, and Overcooked. We also demonstrate our algorithm in a real-world multi-robot setting. Ad hoc teaming (AHT) is an open challenge for multi-agent systems, in which an autonomous agent must successfully coordinate with other unknown agents [1]. Consider a search-and-rescue mission where robots are deployed from different organizations and expected to cooperate with each other on the fly--these robots may have different biases in how they achieve a given objective (e.g., risky vs. risk-averse search) or have different capabilities (e.g., sensing vs. manipulation). Adapting to such differences would enable agents to effectively and autonomously complete tasks where the team is unknown prior to deployment.

Although language model-based chat systems are increasingly used in daily life, most Americans remain non-adopters of chat-based LLMs -- as of June 2025, 66% had never used ChatGPT. At the same time, LLM development and evaluation rely mainly on data from adopters (e.g., logs, preference data), focusing on the needs and tasks for a limited demographic group of adopters in terms of geographic location, education, and gender. In this position paper, we argue that incorporating non-adopter perspectives is essential for developing broadly useful and capable LLMs. We contend that relying on methods that focus primarily on adopters will risk missing a range of tasks and needs prioritized by non-adopters, entrenching inequalities in who benefits from LLMs, and creating oversights in model development and evaluation. To illustrate this claim, we conduct case studies with non-adopters and show: how non-adopter needs diverge from those of current users, how non-adopter needs point us towards novel reasoning tasks, and how to systematically integrate non-adopter needs via human-centered methods.

Multimodal learning systems often struggle in non-stationary environments due to concept drift, where changing data distributions can degrade performance. Modality-specific drifts and the lack of mechanisms for continuous, stable adaptation compound this challenge. This paper introduces LS-OGD, a novel adaptive control framework for robust multimodal learning in the presence of concept drift. LS-OGD uses an online controller that dynamically adjusts the model's learning rate and the fusion weights between different data modalities in response to detected drift and evolving prediction errors. We prove that under bounded drift conditions, the LS-OGD system's prediction error is uniformly ultimately bounded and converges to zero if the drift ceases. Additionally, we demonstrate that the adaptive fusion strategy effectively isolates and mitigates the impact of severe modality-specific drift, thereby ensuring system resilience and fault tolerance. These theoretical guarantees establish a principled foundation for developing reliable and continuously adapting multimodal learning systems.

We introduce HealthDial, a dialogue authoring tool that helps healthcare providers and educators create virtual agents that deliver health education and counseling to patients over multiple conversations. HealthDial leverages large language models (LLMs) to automatically create an initial session-based plan and conversations for each session using text-based patient health education materials as input. Authored dialogue is output in the form of finite state machines for virtual agent delivery so that all content can be validated and no unsafe advice is provided resulting from LLM hallucinations. LLM-drafted dialogue structure and language can be edited by the author in a no-code user interface to ensure validity and optimize clarity and impact. We conducted a feasibility and usability study with counselors and students to test our approach with an authoring task for cancer screening education. Participants used HealthDial and then tested their resulting dialogue by interacting with a 3D-animated virtual agent delivering the dialogue. Through participants' evaluations of the task experience and final dialogues, we show that HealthDial provides a promising first step for counselors to ensure full coverage of their health education materials, while creating understandable and actionable virtual agent dialogue with patients.

Test-time scaling has enabled Large Language Models (LLMs) with remarkable reasoning capabilities, particularly in mathematical domains, through intermediate chain-of-thought (CoT) reasoning before generating final answers. However, the specific sources and mechanisms underlying these reasoning capabilities remain insufficiently understood. Optimization reasoning, i.e. finding extrema under constraints, represents a fundamental abstraction that underpins critical applications in planning, control, resource allocation, and prompt search. To systematically evaluate this capability, we introduce ExtremBench, a benchmark dataset for solving mathematical extremal problems, curated from inequality exercises used for Chinese Mathematical Olympiad and transformed into $93$ standardized extrema-finding problems. We conduct extensive evaluations across various state-of-the-art open-source model families, including the Qwen3, GPT-OSS, and DeepSeek. Our results reveal that LLMs' extremal-solving reasoning capabilities do not always align with those of current mathematical benchmarks such as AIME25 and MATH-500, with some models showing strong general mathematical reasoning but poor extremal-solving skills, and vice versa. This discrepancy highlights a critical gap in current evaluation practices and suggests that existing benchmarks may not comprehensively capture the full spectrum of mathematical reasoning abilities.