method
active
method:latent-anchored-grpo-la-grpoLatent-Anchored GRPO (LA-GRPO)
Token-level auxiliary objective that strengthens optimization of sparse functional tokens during RL by anchoring group-level advantages directly to functional-token positions.
Neighborhood — ranked by edge-count
Frameworks (1)
framework
- Core framework proposing discrete functional tokens as a unified solution for visual reasoning in VLMs, bridging agentic and latent approaches.
Findings (1)
finding
- Gradient Dilution IssuesupportsDuring RL training on ATLAS, sparse functional tokens (2.3% of sequences) receive diluted gradient signals from sequence-level advantages propagated across all tokens.
Methods (1)
method
- RL algorithm used to train the activation verbalizer on open models; samples group of candidate descriptions and applies policy optimization.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Metric measuring the mean MSE between self and other-referencing activations across all hidden MLP/attention layers
- Baseline method using a single orthogonal matrix trained to map source latents to target latents via CL auxiliary loss without behavioral objective.
- Using per-prompt average SAE latent activations and area under precision-recall curve to discriminate aligned from misaligned models
- SAE-derived representation space where individual dimensions activate for distinct concepts, enabling cleaner steering
- Hidden or underdeveloped structures existing 'between the lines' of a configuration that can be enhanced and developed through harmony-seeking computation.
- Entities that become visible as centers in a configuration (e.g., rectangles of white space around a dot) that were not present before.
- Statistical regularities stored in pretrained models.
- Cost-efficient training algorithm used by DeepSeek-R1 for RL-based reasoning