TRAM: Enhancing Multimodal Reasoning with Trajectory-Derived Auxiliary Memory
2026-08-03 • Computation and Language
Computation and Language
AI summaryⓘ
The authors study how large AI models that work with both images and text can sometimes forget important parts of their reasoning when solving complex problems. They found that just focusing on the image isn’t enough; the model also needs to remember the intermediate steps and conclusions it makes along the way. To fix this, they created TRAM, a method that helps the model store and use its own reasoning process better without extra training. Testing TRAM on various tasks showed it helps models reason more accurately in math, science, and visual problems.
Multimodal Large Reasoning Modelsvisual groundingreasoning trajectoryauxiliary memorydecoder layerslatent memoryresidual pathwaymulti-step inference
Authors
Kang Liu, Zijing Wang, Yongkang Liu, Mengjie Zhao, Xiaocui Yang, Shi Feng, Yifei Zhang, Daling Wang
Abstract
Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.