arXiv:2606.19828v2 Announce Type: replace Abstract: 3D multimodal large language models (3D MLLMs) describe a 3D object as a whole but cannot address, name, or reason about its parts. Prior part-awar...
A*-Thought-V2 is a framework that models Chain-of-Thought reasoning as a geometric trajectory in a 3D PCA space, using explicit-implicit latent tokens to compress steps that deviate from the main question-to-solution direction. The method measures alignment angles to decide which steps remain text and which become latent, and introduces stepwise embedding forcing and label forcing to train the architecture. Experiments on Qwen models show up to 2.6% accuracy gains, halved response length, and significant reductions in computation and training time.
arXiv:2605.16893v2 Announce Type: replace Abstract: Recent studies have introduced memory-augmented architectures that decouple knowledge storage from neural computation, enabling more direct access...
Subspace-Decomposed JEPAs (SD-JEPA) split the latent space of Joint-Embedding Predictive Architectures into two orthogonal subspaces: a low-dimensional progression subspace trained with a cosine-margin triplet loss and a high-dimensional content subspace regularised by SIGReg. The authors prove that the anti-collapse forces act on disjoint coordinates, allowing additive composition rather than competition. SD-JEPA outperforms the LeWM baseline on most control benchmarks and the strongest non-LeWM JEPA baseline on Push‑T, with a subspace-ablation confirming the split as essential. The 1‑D angular progression coordinate serves as a scene-aware compass, advancing with task progress, regressing on backtracking, and relocalising under perturbations to separate surprise from meaning.
The paper introduces DARTS, a method for tuning decoder representations during model merging. It addresses representation bias in autoregressive decoders by using an entropy‑weighted L1 loss and a per‑position additive bias to correct errors that accumulate across token positions. Experiments on code generation, mathematical reasoning, and instruction following with Llama‑2‑7B show that DARTS improves performance over standard surgery while adding only 0.1% extra parameters.
arXiv:2608.21447v1 Announce Type: new Abstract: Structured-language models such as SceneScript reconstruct a scene as a short program of parametric commands, an inherently editable and semantically e...
arXiv:2511. 08577v3 Announce Type: replace-cross Abstract: Improving the reasoning abilities of Large Language Models (LLMs), especially under parameter constraints, is crucial for real-world applications.
arXiv:2607. 01792v1 Announce Type: cross Abstract: While decoder-only LLMs excel at a vast array of natural language tasks, it suffers from an asymmetric information flow induced by causal attention: later tokens are richer in contextual grounding than earlier ones.
arXiv:2610.20592v1 Announce Type: new Abstract: Existing geometry-aware Multimodal Large Language Models (MLLMs) increasingly rely on external 3D encoders to provide structural priors, but executing...
arXiv:2608. 06111v1 Announce Type: cross Abstract: Positional embeddings (PE) in Transformers encode token distance and order but are largely agnostic to \textit{syntactic structure}.
The paper investigates how verb–noun decomposition, a common strategy for recognizing assembly actions, generalizes to novel combinations of familiar components. Through a systematic study on three datasets (MECCANO, HAViD, and IMPACT), the authors find that while decomposition avoids the zero‑probability ceiling of atomic classifiers, its performance still heavily depends on the co‑occurrence patterns seen during training. The analysis reveals that errors concentrate on the larger‑vocabulary component, that shared‑encoder training can entangle components and worsen generalization, and that these issues stem from primitive support, vocabulary asymmetry, and component entanglement.
The paper introduces PFIR, a Part-level Functional Inference for Recognition framework that reverses the typical function-as-annotation flow in 3D multimodal large language models (MLLMs). By leveraging PartNeXt segmentation and a curriculum of 507K QA pairs with functional chains of thought, PFIR trains a dual-objective reinforcement learning model (GRPO) to perform strict functional reasoning and descriptive grounding. In zero‑shot tests, PFIR improves accuracy from 60.5% to 66.7%, especially on geometrically ambiguous samples, demonstrating that part‑level functional reasoning enhances generalizable 3D recognition.