Federated inference, in the form of one-shot federated learning, edge ensembles, or federated ensembles, has emerged as an attractive solution to combine predictions from multiple models. This paradigm enables each model to remain local and proprietary while a central server queries them and aggregates predictions. Yet, the robustness of federated inference has been largely neglected, leaving them vulnerable to even simple attacks. To address this critical gap, we formalize the problem of robust federated inference and provide the first robustness analysis of this class of methods. Our analysis of averaging-based aggregators shows that the error of the aggregator is small either when the dissimilarity between honest responses is small or the margin between the two most probable classes is large. Moving beyond linear averaging, we show that problem of robust federated inference with non-linear aggregators can be cast as an adversarial machine learning problem. We then introduce an advanced technique using the DeepSet aggregation model, proposing a novel composition of adversarial training and test-time robust aggregation to robustify non-linear aggregators. Our composition yields significant improvements, surpassing existing robust aggregation methods by 4.7 - 22.2% in accuracy points across diverse benchmarks.

Achieving stable and robust perceptive locomotion for bipedal robots in unstructured outdoor environments remains a critical challenge due to complex terrain geometry and susceptibility to external disturbances. In this work, we propose a novel reward design inspired by the Linear Inverted Pendulum Model (LIPM) to enable perceptive and stable locomotion in the wild. The LIPM provides theoretical guidance for dynamic balance by regulating the center of mass (CoM) height and the torso orientation. These are key factors for terrain-aware locomotion, as they help ensure a stable viewpoint for the robot's camera. Building on this insight, we design a reward function that promotes balance and dynamic stability while encouraging accurate CoM trajectory tracking. To adaptively trade off between velocity tracking and stability, we leverage the Reward Fusion Module (RFM) approach that prioritizes stability when needed. A double-critic architecture is adopted to separately evaluate stability and locomotion objectives, improving training efficiency and robustness. We validate our approach through extensive experiments on a bipedal robot in both simulation and real-world outdoor environments. The results demonstrate superior terrain adaptability, disturbance rejection, and consistent performance across a wide range of speeds and perceptual conditions.

Sparse Autoencoders (SAEs) can extract interpretable features from large language models (LLMs) without supervision. However, their effectiveness in downstream steering tasks is limited by the requirement for contrastive datasets or large activation storage. To address these limitations, we propose CorrSteer, which selects features by correlating sample correctness with SAE activations from generated tokens at inference time. This approach uses only inference-time activations to extract more relevant features, thereby reducing spurious correlations. It also obtains steering coefficients from average activations, automating the entire pipeline. Our method shows improved task performance on QA, bias mitigation, jailbreaking prevention, and reasoning benchmarks on Gemma-2 2B and LLaMA-3.1 8B, notably achieving a +3.3% improvement in MMLU performance with 4000 samples and a +27.2% improvement in HarmBench with only 108 samples. Selected features demonstrate semantically meaningful patterns aligned with each task's requirements, revealing the underlying capabilities that drive performance. Our work establishes correlation-based selection as an effective and scalable approach for automated SAE steering across language model applications.

We address the task of zero-shot video classification for extremely fine-grained actions (e.g., Windmill Dunk in basketball), where no video examples or temporal annotations are available for unseen classes. While image-language models (e.g., CLIP, SigLIP) show strong open-set recognition, they lack temporal modeling needed for video understanding. We propose ActAlign, a truly zero-shot, training-free method that formulates video classification as a sequence alignment problem, preserving the generalization strength of pretrained image-language models. For each class, a large language model (LLM) generates an ordered sequence of sub-actions, which we align with video frames using Dynamic Time Warping (DTW) in a shared embedding space. Without any video-text supervision or fine-tuning, ActAlign achieves 30.5% accuracy on ActionAtlas--the most diverse benchmark of fine-grained actions across multiple sports--where human performance is only 61.6%. ActAlign outperforms billion-parameter video-language models while using 8x fewer parameters. Our approach is model-agnostic and domain-general, demonstrating that structured language priors combined with classical alignment methods can unlock the open-set recognition potential of image-language models for fine-grained video understanding.

Large language models (LLMs) are increasingly evaluated on single-answer multiple-choice tasks, yet many real-world problems require identifying all correct answers from a set of options. This capability remains underexplored. We introduce SATA-BENCH, the first dedicated benchmark for evaluating LLMs on Select All That Apply (SATA) questions across diverse domains, including reading comprehension, law, and biomedicine. Our evaluation of 27 open-source and proprietary models reveals a significant gap: even the strongest model achieves only 41.8% exact match, exposing LLMs' inability to reliably identify all correct answers. We find that this weakness stems from two core challenges: selection bias - models favor certain choices regardless of content, and count bias - models fail to predict the correct number of answers. To address these issues, we propose Choice Funnel, a decoding strategy that combines token debiasing with adaptive thresholding to guide models toward complete and accurate selections. Choice Funnel achieves up to 29% higher exact match than competitive baselines while reducing inference cost by over 64%. Our findings expose fundamental limitations in current LLMs and introduce a new framework for diagnosing and improving multi-answer reasoning. We release SATA-BENCH and Choice Funnel to promote LLM development for robust decision-making in realistic, multi-answer applications.

Large Language Models (LLMs) represent substantial intellectual and economic investments, yet their effectiveness can inadvertently facilitate model imitation via knowledge distillation (KD). In practical scenarios, competitors can distill proprietary LLM capabilities by simply observing publicly accessible outputs, akin to reverse-engineering a complex performance by observation alone. Existing protective methods like watermarking only identify imitation post-hoc, while other defenses assume the student model mimics the teacher's internal logits, rendering them ineffective against distillation purely from observed output text. This paper confronts the challenge of actively protecting LLMs within the realistic constraints of API-based access. We introduce an effective and efficient Defensive Output Generation (DOGe) strategy that subtly modifies the output behavior of an LLM. Its outputs are accurate and useful for legitimate users, yet are designed to be misleading for distillation, significantly undermining imitation attempts. We achieve this by fine-tuning only the final linear layer of the teacher LLM with an adversarial loss. This targeted training approach anticipates and disrupts distillation attempts during inference time. Our experiments show that, while preserving the performance of the teacher model, student models distilled from the defensively generated outputs demonstrate catastrophically reduced performance, demonstrating DOGe as a practical safeguard against KD-based model imitation.

Mathematical reasoning is central to artificial intelligence, with applications in education, code generation, and research-level mathematical discovery. Mathematical competitions highlight two problem types: theorem proving, requiring rigorous proofs, and answer construction, requiring creative generation and formal verification of mathematical objects. Existing research reveals that LLMs can tackle difficult answer-construction tasks but are prone to errors from hallucinations and unverifiable steps, while symbolic methods guarantee rigor but falter in creative answer construction. This raises a key understudied question: how to solve answer-construction problems while preserving both LLM creativity and mathematical rigor? To address this problem, we introduce the Enumerate-Conjecture-Prove (ECP) framework, a modular neuro-symbolic method integrating LLM-based enumeration and pattern-driven conjecturing with formal theorem proving in Lean, and ConstructiveBench, a dataset of 3,640 formal answer-construction problems from math competitions. ECP is model agnostic and shows consistent improvements over pure LLM baselines: on the subset of PutnamBench for answer construction, ECP formally solves 6 out of 337 answer-construction problems end to end (up from 4 without ECP) using GPT-5 mini and DeepSeek-Prover-V2-7B. On ConstructiveBench, ECP achieves 33.1% end-to-end state-of-the-art accuracy (up from 32.5%), demonstrating its potential to advance formal mathematical reasoning by combining LLM conjecturing with formal verification. Our code and dataset are publicly available at GitHub (https://github.com/sunjia72/ECP) and Hugging Face (https://huggingface.co/datasets/sunjia72/ConstructiveBench).

Large Language Models (LLMs) have demonstrated strong generative capabilities but remain prone to inconsistencies and hallucinations. We introduce Peer Elicitation Games (PEG), a training-free, game-theoretic framework for aligning LLMs through a peer elicitation mechanism involving a generator and multiple discriminators instantiated from distinct base models. Discriminators interact in a peer evaluation setting, where utilities are computed using a determinant-based mutual information score that provably incentivizes truthful reporting without requiring ground-truth labels. We establish theoretical guarantees showing that each agent, via online learning, achieves sublinear regret in the sense their cumulative performance approaches that of the best fixed truthful strategy in hindsight. Moreover, we prove last-iterate convergence to a truthful Nash equilibrium, ensuring that the actual policies used by agents converge to stable and truthful behavior over time. Empirical evaluations across multiple benchmarks demonstrate significant improvements in factual accuracy. These results position PEG as a practical approach for eliciting truthful behavior from LLMs without supervision or fine-tuning.

Sim2Real aims at training policies in high-fidelity simulation environments and effectively transferring them to the real world. Despite the developments of accurate simulators and Sim2Real RL approaches, the policies trained purely in simulation often suffer significant performance drops when deployed in real environments. This drop is referred to as the Sim2Real performance gap. Current Sim2Real RL methods optimize the simulator accuracy and variability as proxies for real-world performance. However, these metrics do not necessarily correlate with the real-world performance of the policy as established theoretically and empirically in the literature. We propose a novel framework to address this issue by directly adapting the simulator parameters based on real-world performance. We frame this problem as a bi-level RL framework: the inner-level RL trains a policy purely in simulation, and the outer-level RL adapts the simulation model and in-sim reward parameters to maximize real-world performance of the in-sim policy. We derive and validate in simple examples the mathematical tools needed to develop bi-level RL algorithms that close the Sim2Real performance gap.

Dialogue plays a crucial role in educational settings, yet existing evaluation methods for educational applications of large language models (LLMs) primarily focus on technical performance or learning outcomes, often neglecting attention to learner - LLM int eractions. To narrow this gap, this AIED Doctoral Consortium paper presents an ongoing study employing a dialogue analysis approach to identify effective pedagogical strategies from learner - LLM dialogues. The proposed approach involves dialogue d ata collection, dialogue act (DA) annotation, DA pattern mining, and predictive model building. Early insights are outlined as an initial step toward future research. The work underscores the need to evaluate LLM - based educational applications by focusing on dialogue dynamics and pedagogical strategies.