Gradient Dilution Issue

During RL training on ATLAS, sparse functional tokens (2.3% of sequences) receive diluted gradient signals from sequence-level advantages propagated across all tokens.

Source paper

extracted_from

ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both

Ziyu Guo · Rain Liu · Xinyan Chen · Pheng-Ann Heng

Neighborhood — ranked by edge-count

Claims (1)

claim

Token-level supervision enables models to learn functional-token invocation from reasoning context
contradicts
ATLAS author's assertion that functional tokens optimized via standard cross-entropy loss learn when and how to invoke operations from surrounding text.

Communities (3)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Mechanistic structure of transformer attention computations
members_of
Identifies distributed algorithms implemented across attention heads, with focus on causal masking limitations and emergent capabilities via activation manifold steering.
Functional tokens for emergent model reasoning
members_of
Unsupervised learning of interpretable task tokens through gradient flow and vocabulary constraints, enabling reasoning without visual supervision.

Methods (1)

method

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

Related by similarity (8)

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.

Gradient Descentmethod0.788
Used for updating hidden state expectations; provides dynamical process theory testable against neuronal data
Gradient conflictconcept0.784
When gradients of different tasks have negative cosine similarity, harming multi-task learning.
Gradient methodmethod0.775
Optimization technique that computes weight changes by following the gradient of an error function; contrasted with evolutionary stochastic search.
Gradientsconcept0.763
The property that qualities vary slowly, subtly, gradually across the extent of each living thing; gradients arise as natural responses to changing circumstances and create field-like character that points toward and establishes centers
Gradients are nearly non-existent in the modern environment because naive standardization, mass-production, and regulation of sizes all work against their formationclaim0.756
Claim that one of the most powerful forms of life has been almost removed from the environment by industrial production norms
gradient-magnitude balancingconcept0.747
Addressing disparity in gradient magnitudes across tasks at the gradient level
Gradient-based data attributionmethod0.743
Baseline method against which probe-based ranking is compared; more computationally expensive.
Gradients Transformationconcept0.736
A structure-preserving transformation: using gradual change across space to soften and intensify transitions.