finding
active
finding:length-normalization-prevents-degenerate-tool-calling-trajectories-and-repeated-tool-calls-without-normalizationLength normalization prevents degenerate tool-calling trajectories and repeated tool calls without normalization.
Empirical result showing that without length normalization, RL training produces rapidly increasing tool usage with performance collapse and repetitive tool calls.
Source paper
extracted_from(2025) · Xuan-Phi Nguyen · Shrey Pandit · Revanth Gangi Reddy · Aimin Xu +3
Neighborhood — ranked by edge-count
Communities (2)
community
- Few-shot anchoring & latent structuremembers_ofHow minimal examples disambiguate and recruit latent arithmetic/reasoning interpretations in LLMs
- Unified Competency Control Theory (UCCT)members_ofFormal framework modeling prompt/context design as latent competency toggling via anchor budget regularization, with measurable quantities ρd, dr, k, S enabling cross-domain diagnostics.
Methods (1)
method
- Novel modification to REINFORCE that normalizes step-level advantage by trajectory length to prevent long but low-quality trajectories from dominating training.
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.
- Theoretical interpretation.
- Fundamental design principle driving Oberon: proper abstraction enables modular structure and maintainability.
- Subclaim of how MCA speeds evolution.
- EI and normalized EI could serve as a unified metric for out-of-distribution generalization.claim0.716Conjecture that maximizing EI yields causal representations invariant to distribution shifts.
- Clarification that the multi-axis delegation is sequential and deterministic.
- Evolution learns to generalize beyond default morphologies, producing problem-solving machines.claim0.710Argues that evolutionary learning goes beyond specific adaptations.