@HaiyuWu1: Locally straightening the latent trajectory can largely reduce the optimization difficulty of world model planning! Sou…
Summary
The paper proposes locally straightening latent trajectories by maximizing cosine similarity between adjacent velocity vectors to reduce optimization difficulty in world model planning, with experiments on PushT dataset achieving 36% success rate for long-horizon planning.
View Cached Full Text
Cached at: 07/15/26, 03:42 AM
Locally straightening the latent trajectory can largely reduce the optimization difficulty of world model planning!
Sounds confusing? Not sure how to achieve it? Don’t worry.
After the math derivation, the final implementation is very simple: increasing the cosine similarity between adjacent velocity vectors.
L_curv = 1 - cos(v_t, v_(t+1))
Another interesting finding: Applying this loss to patch tokens is more effective than applying it to the CLS token (we are far from fully leveraging the pre-trained model, lol).
Since this paper was done before Le-WM, I wonder if anyone has tested this loss on Le-WM? In this paper, long-horizon planning remains quite challenging (i.e., planning horizon: 50, dataset: PushT, best success rate: 36.00%).
#WorldModel
Similar Articles
The Objective Is the Bottleneck: Latent World Models Encode What Their Planners Cannot Use
The paper investigates why latent world models fail at long-horizon planning and finds the bottleneck is the planning objective (squared latent distance), not the predictor's accuracy; replacing the objective with a learned cost dramatically improves planning performance.
Topology-Enhanced Alignment for Large Language Models: Trajectory Topology Loss and Topological Preference Optimization
This paper introduces a topology-enhanced alignment framework for LLMs, utilizing trajectory topology loss and topological preference optimization based on persistent homology to regularize semantic trajectories in hidden space.
Investigating Implicit Latent Trajectory Shifts: Bypassing Alignment via Long-Form Coherent Context
An empirical study investigating how long, semantically dense benign text can shift a model's latent space trajectory, diluting initial system prompts and bypassing post-training alignment constraints, as observed in both closed and open-source models.
@swyx: very notable trajectory comparison writeup here buried in the RLM paper from @a1zhang and @lateinteraction. an open sec…
Discusses a paper by Alex Zhang and Omar that reveals how frontier models can cheat on benchmarks by training on test lookalikes, and proposes using NLP distance metrics on hidden trajectories to detect such cheating.
TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment
This paper introduces Trajectory Matching Policy Optimization (TMPO), a method for aligning diffusion models that addresses reward hacking and visual mode collapse by matching trajectory-level reward distributions rather than maximizing scalar rewards.