Skip to content
AI-generated research idea🟠Medium💡Moderate InnovationRelated work found

Infinite-Context Linear State Generalization for Feedforward Reconstruction

Replace standard test-time training memory updates with infinite-context linear state-space models to eliminate error accumulation and trajectory drift during long sequence reconstruction.

Time

4 days

Compute level

Large cluster

Resources

32 NVIDIA H200 GPUs for 4 days

What the paper used

32 NVIDIA H100 GPUs followed by 32 NVIDIA H200 GPUs

Prerequisites

transformer architectures and state-space models

Scope

new model architecture

The Gap

The TTT module's fast weights are constrained by the training sequence length (e.g., 128 frames or 20 chunks). During inference on very long sequences, the recurrent TTT formulation struggles to generalize beyond this trained chunk count, causing error accumulation and trajectory drift that currently forces the use of hard periodic state resets (every five windows), which destroys historical global context.

“While TTT can compress and propagate information for long contexts with linear complexity, its compressed memory is inherently lossy.”
From the paper, Section 4.1, page 4. Matched against the PDF text.

Limitation

The TTT memory module fails to generalize to unbounded sequence lengths at inference time, necessitating destructive state resets that sacrifice long-term global context.

Proposed Solution

Replace the standard TTT memory update with an advanced infinite-context linear sequence model (such as Mamba-2 or a dynamically decaying state-space model) combined with a soft-reset or hierarchical compression mechanism. This would allow the model to retain sparse global anchor points indefinitely rather than completely clearing the hidden state.

The paper’s authors list this direction as future work, so they may already be pursuing it.

Why This Matters

Solving this will eliminate the need for manual state resets, enabling truly infinite-context 3D reconstruction and eliminating trajectory drift in unbounded real-world video streams.

How to Get Started

How to test it: We hypothesize that replacing standard TTT updates with an infinite-context linear sequence model and soft-reset mechanisms improves long-sequence 3D reconstruction. We will compare against standard LoGeR on the VBR dataset using Absolute Trajectory Error (ATE), stopping if ATE increases by more than 5% on sequences exceeding 10k frames.

  1. Analyze the existing TTT memory update bottleneck and the impact of hard periodic state resets on sequence reconstruction.
  2. Integrate an infinite-context linear sequence model, such as Mamba-2, into the feedforward reconstruction architecture.
  3. Develop a soft-reset or hierarchical compression mechanism to retain sparse global anchor points without clearing hidden states.
  4. Evaluate the modified reconstruction pipeline on long-horizon sequence datasets measuring drift and reconstruction fidelity.

Code linked to this paper on Hugging Face: github.com/Junyi42/LoGeR

Suggested by AI. Adapt it to your setup.

Has This Been Done?

Related work found

Papers like Mamba-VGGT integrate Mamba memory mechanisms for long-sequence video geometry, and Latent Context Compilation distills long contexts into compact portable memory. However, none of the listed works specifically replace TTT updates with an infinite-context linear state model utilizing soft-reset and hierarchical compression for feedforward reconstruction.

We searched OpenAlex and arXiv for “Replace standard test-time training memory updates with infinite-context linear state-space models to eliminate error accumulation and trajectory drift during long sequence reconstruction.” and “Infinite-Context Linear State Generalization for Feedforward Reconstruction” and an AI compared the idea with the top 10 results. A small search can miss relevant work.

Source Paper

LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory

Junyi Zhang, Charles Herrmann, Junhwa Hur +5 more · 2026

This paper introduces LoGeR, a novel feedforward geometric foundation model designed for long-context dense 3D reconstruction. It overcomes the quadratic complexity of global attention and sequence length limitations by utilizing a chunk-wise processing architecture with a hybrid memory system, achieving state-of-the-art results on extremely long video sequences.

Related Ideas

Is this a good idea?

Share onXLinkedIn