ATLAS: Agentic or Latent Visual Reasoning

Core framework proposing discrete functional tokens as a unified solution for visual reasoning in VLMs, bridging agentic and latent approaches.

Neighborhood — ranked by edge-count

thinker

method

Latent-Anchored GRPO (LA-GRPO)
implements
Token-level auxiliary objective that strengthens optimization of sparse functional tokens during RL by anchoring group-level advantages directly to functional-token positions.

concept

Qwen2.5-VL-7B
implements
Base vision-language model used to instantiate ATLAS.
Agentic Visual Reasoning
associated_with
Paradigm where VLM acts as controller generating code or tool calls to external modules for visual operations, incurring context-switching latency.
Functional Token
implements
A discrete token in the vocabulary that represents a visual operation (e.g., <|Line|>, <|Shape|>, <|Text|>), generated via next-token prediction within autoregressive sequences.
Unified Models Generate Intermediate Images
associated_with
Paradigm where VLMs explicitly generate pixel-level intermediate images for visual reasoning, incurring high computational overhead.

dataset

ATLAS-178K
cites
Curated SFT dataset with 178K examples covering 40+ visual reasoning tasks, annotated with functional-token trajectories to provide supervised training signal.
BLINK
cites
Evaluation benchmark for visual reasoning tasks used to assess ATLAS performance.
V*
cites
Evaluation benchmark for visual reasoning used to assess ATLAS performance.
WeMath
cites
Evaluation benchmark for mathematical visual reasoning used to assess ATLAS performance.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

ATLAS combines the strengths of agentic and latent reasoning by using a single discrete word (functional token) that serves both roles.claim0.865
Core claim of the paper: a unified token bridges the gaps.
One word is enough for both agentic and latent visual reasoning.claim0.807
The paper's thesis from the title.
Vector-based navigation using grid-like representations in artificial agents (Banino et al., 2018)concept0.768
Demonstrated grid cell emergence in RNNs trained on spatial navigation; related work category 4.
latent reasoningconcept0.766
Reasoning approach using learnable hidden embeddings.
agentic reasoningconcept0.764
Reasoning approach using code or tool calls executed by an agent.
Recent alternatives include agentic reasoning through code or tool calls, and latent reasoning with learnable hidden embeddings.claim0.762
Statement of existing alternatives to direct generation.
visual reasoningconcept0.757
Visual reasoning tasks often interleaved with intermediate visual states; promising direction in the field.
One word is enough for both agentic operation and latent reasoning unitconcept0.752
Core idea of ATLAS: a single discrete token serves dual purpose of operation specification and latent reasoning.