Deqian Kong

understanding how networks represent, reason, and learn

Research

My work follows three questions. For each I give the idea in a sentence or two and the papers that carry it. The complete record is under Publications.

Neural representations and computation

What the cells of a brain, or the units of a network, represent, and what that representation makes easy to compute.

Place Cells as Multi-Scale Position Embeddings. NeurIPS 2025

The population of place cells forms an embedding of position whose inner products are normalized random-walk transition probabilities. In this embedding, planning follows a straight path.

\langle\, h(x,t),\; h(y,t) \,\rangle \;=\; q(y \mid x,\, t) (1)
Multi-scale place fields and planned paths in open fields and mazes

Figure 1. Place fields at several time scales, and the paths planned from them in open fields and mazes.

Latent Thought Models with Variational Bayes Inference-Time Computation. ICML 2025

A language model with explicit thought vectors, inferred for each sequence by fast learning of local parameters, which complements the slow learning of global ones.

Architecture of the Latent Thought Model, and validation perplexity against inference steps

Figure 2. Left: thought vectors z condition every layer through cross-attention. Right: validation perplexity against inference steps.

Learning, reasoning, and planning

How a model learns from experience, and how it uses what it has learned to think ahead.

Learning to Plan from Random Exploration. Preprint, 2026

Planning can be based on random exploration directly, without policy optimization.

Latent Plan Transformer for Trajectory Abstraction. NeurIPS 2024

A latent plan controls a whole trajectory. Only the whole plan is evaluated, without stepwise rewards, which avoids the credit assignment problem.

Maze trajectories from the dataset and generated by the Latent Plan Transformer

Figure 3. Trajectories in two mazes: from the dataset, and generated from a latent plan.

Probabilistic modeling and scientific discovery

The same probabilistic tools, applied to molecules and to many-body physics.

Molecule Design by Latent Prompt Transformer. NeurIPS 2024 Spotlight

A continuous latent design is easier to optimize than a search over the discrete space of molecules.

Designed molecules with binding affinities, and docked poses

Figure 4. Designed molecules with their binding affinities (a), and docked to the protein (b).

FFT-Accelerated Auxiliary Variable MCMC for Fermionic Lattice Models. APS March Meeting 2026

A determinant-free sampler with O(N \log N) complexity.

All publications →