281

Papers

10/10

Roadblocks Active

0

Connections

Mellum 2 is a Mixture-of-Experts language model with 64 total experts but only 8 active per token, letting it match the quality of 4B-14B dense models while running at the cost of a 2.5B one. Its training curriculum progressively increases the share of code from 23% to 59% over 10.6 trillion tokens, and a Multi-Token Prediction head doubles as a draft model for faster inference. This matters because it demonstrates a concrete, reproducible recipe for achieving high code and math capability without proportionally scaling compute costs.

██████████ 0.9 efficiency-scaling Preprint

ReadSaveConnections

LongTraceRL trains language models to reason over very long documents by using a search agent's browsing history to build training examples — documents the agent read but did not cite become deliberate distractors, making the task harder and more realistic. A fine-grained reward signal based on which specific reasoning-chain entities appear in correct answers helps distinguish between responses that are correct for the right versus wrong reasons. The method outperforms strong baselines on five long-context benchmarks across models ranging from 4B to 30B parameters, showing that how training contexts are constructed matters as much as the reward signal itself.

██████████ 0.9 long-context Preprint

ReadSaveConnections

When shown gender-ambiguous images with minimal prompting, vision-language models consistently output male gender — even for occupations strongly associated with women — despite their internal activations encoding female associations mid-network. Layer-wise analysis using a new metric called LALS reveals that female signals peak internally around the middle of the network and are suppressed before generation, while male signals amplify end-to-end. This decoupling between what a model internally represents and what it outputs is a meaningful safety concern because it means standard output audits would underestimate the bias present inside the model.

██████████ 0.9 alignment-safety Preprint

ReadSaveConnections

SpatialAct tests whether vision-language models can not only answer spatial questions but act on them — issuing move, rotate, and scale commands inside a 3D simulator and updating their beliefs as the environment changes. Current VLMs perform well on one-shot spatial questions but fail sharply when required to refine their actions over multiple turns based on feedback, a gap that human participants do not show. With 4,355 question-action pairs across 333 scenes, the benchmark separates static spatial knowledge from the dynamic spatial reasoning needed for real-world agents.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

ERGeoBench evaluates whether multimodal models can determine their location on Earth from street-view imagery, progressing from a single photo to a full panorama to an interactive mode where the model can pan, tilt, and zoom before answering. Across 2,207 globally distributed scenes and nine models, current systems can infer broad geographic context but fail at precise metric localization and at maintaining consistent spatial beliefs across sequential views. The benchmark links geo-localization failure directly to weaknesses in foundational perception and spatial awareness, providing a concrete diagnostic ladder for where models break down.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

nuReasoning provides 20,000 real-world driving clips annotated with spatial, decision, and counterfactual reasoning questions specifically targeting rare and difficult scenarios where standard perception alone is insufficient. Fine-tuning vision-language models on this data improves driving-specific question answering, and — notably — adding reasoning supervision to action models improves their planning even when the textual reasoning output is discarded at inference time. This suggests that reasoning annotations act as a structuring force on learned representations, not just as a surface output.

██████████ 0.9 reasoning-reliability Preprint

ReadSaveConnections

When multiple AI agents share a common workspace to collaborate on visual reasoning tasks, unverified notes left by one agent get picked up as evidence by subsequent agents — a failure mode the authors call Noise Reinforcement — while the added context also pushes models toward vague, short answers (Policy Collapse). These failure modes were identified using the CoSee auditing framework across three document visual QA benchmarks under single-GPU resource constraints. The finding is important because shared workspaces are a popular design pattern for multi-agent systems, and this work shows they can actively worsen accuracy rather than improve it.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

FBHM exposes that state-of-the-art vision-language models drop from high accuracy on standard hateful meme datasets to near-random performance on a new 5,000-meme benchmark that systematically varies the target community while holding the rhetorical mechanism constant. This collapse reveals that models learn dataset-specific shortcuts — keying on who is targeted rather than how hate is expressed — rather than genuine cross-modal reasoning. A lightweight intervention method using learnable steering vectors (LSV) with only 500 examples is proposed as a more efficient fix than full fine-tuning.

██████████ 0.9 data-quality-curation Preprint

ReadSaveConnections

TouchSafeBench tests whether vision-language models can detect when a robot is about to collide with a person or a scene object during navigation and rearrangement tasks, using physics-engine contact signals from the Habitat 3.0 simulator as ground truth. The best-performing models achieve a Macro-F1 below 50%, and providing explicit depth information does not reliably help models infer collision risk — suggesting VLMs lack the grounding needed for basic physical safety judgments. Robot-scene contact is harder to classify than human-contact risk, pointing to a specific gap in how these models understand rigid-body geometry.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

DeMaVLA adapts a Qwen3-VL vision-language backbone into a robot action model by pruning every other transformer layer to create a lightweight action expert, then training it with flow matching on roughly 5,000 hours of dual-arm manipulation demonstrations augmented by corrective trajectories collected when the robot failed. The pruning approach preserves alignment with the VLM backbone while reducing compute, and mixing failure-recovery data across clothing categories improves generalization beyond category-specific policies. Results on both the RoboTwin simulation benchmark and real household folding tasks suggest this is a practical path toward general-purpose deformable object manipulation.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections