OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

Study of Human Push Recovery

Walking and push recovery controllers for humanoid robots have advanced throughout the years while unsolved gaps leading to undesirable behaviours still exist. Because previous studies are mainly pure engineering methods, while the use of data-driven methods has made impressive achievements in the field of control, we set motivation for exploration of control laws applied by human beings. Successful findings may help fill the gaps in current engineering-based controllers and can potentially improve the performance. In this thesis, we show our complete design and implementation of a set of experiments to collect and process human data, as well as data analysis for model fitting. Using the processed motion data and force data collected, we export the position, velocity and acceleration of our participants' centres of mass to form a one-dimensional point mass model defined by the direction of pushes. As a result, we find that proportional-derivative (PD) control can describe the underlying control law that people used and different PD gains are set for different phases of a push recovery trial within the scope of our study. Our final PD control fittings have an average root mean square error below 0.1m with above 90% of a trial's data points are taken into account on average. We also explore how different error metrics that the PD control model uses influence its performance but neither of the two metrics we proposed can help improve the performance. Finally, we have statistics on how people switch push recovery strategies based on their centre of mass properties at the start of the push. We find that the further the centre of mass is from the steady-state point and the higher the velocity it has, a person is more likely to make a step for push force compensation.