The paper investigates where the transition from memorization to generalization occurs within a Transformer. Using Transition Games—activation games with paired non-generalizing controls—it finds that distributed utility gain is driven by a block‑0 attention bias, with degree‑two modes explaining 67–92% of the contrast across replacement games. A detailed path study shows that block‑1 MLP mediates more of this effect than other downstream paths, contradicting the common "MLP memorizes, attention generalizes" hypothesis and suggesting that grokking is a spectral recoding of an existing distributed circuit rather than a module switch.
arXiv:2607. 05104v1 Announce Type: cross Abstract: Grokking -- the delayed onset of generalization long after a network has fit its training set - -is usually studied in models too large to read completely and reported from single training runs.
arXiv:2607. 26648v1 Announce Type: cross Abstract: Spiking neural networks (SNNs) are promoted as an energy-efficient substrate because sparse, event-driven activity replaces dense multiply-accumulates with cheap accumulates.
arXiv:2607. 08032v1 Announce Type: new Abstract: Large language models, and the agents built on them, spend an ever-growing share of their compute and memory on remembering: caching attention keys and values, carrying long prompts, maintaining recurrent state, and storing what happened in previous turns and sessions.
arXiv:2607. 04333v1 Announce Type: new Abstract: Grokking -- generalization arriving long after training-set interpolation -- can be accelerated by structure-agnostic interventions: gradient filtering, weight-norm clamping, geometric penalties on hidden representations.
arXiv:2606. 08105v1 Announce Type: new Abstract: When attention concentrates on a single token, a sink, what is the model actually computing?
Spiking neural networks (SNNs) are promoted as an energy-efficient substrate because sparse, event-driven activity replaces dense multiply-accumulates with cheap accumulates. We argue the energy dividend of sparsity is not a property of SNNs but of the task.
Grokking -- where a transformer on modular arithmetic suddenly transitions from near-chance to near-perfect validation accuracy -- is attributed to a Fourier circuit, but its timing, causal structure, and controllability remain poorly understood. We introduce the Frequency Synchronization Degree (FSD), a normalised, permutation-tested metric for Fourier circuit synchronisation requiring no prior circuit knowledge.
The paper introduces Recuris, a recursive Experiential‑Working Memory architecture that lets long‑horizon agents track task progress and select skills based on current needs rather than full history. By coupling working memory with experiential memory, execution becomes structured evidence that localizes failures to specific memory components, enabling a bounded recursive memory‑evolution loop. Across four benchmarks and ten models, Recuris improves task success in 35 of 37 model‑benchmark pairs, raising state‑of‑the‑art performance on tau‑bench and SkillFlow and reducing common long‑horizon failures by up to 80%.
Learned KV-cache eviction often faces a soft-to-hard mismatch: during training, differentiable gates typically attenuate token contributions, whereas inference saves memory only when KV entries are ph...
The paper introduces a Mixture‑Gated Theory of Anisotropic Representation Dynamics that decomposes Transformer self‑attention into value transport, query gating, and key gating. It shows that the geometry of attended representations can cause a representation difference to contract or expand, a phenomenon termed gating‑induced regime switching. The authors derive a spectral criterion to predict switching, quantify its strength, and identify subspaces where switching occurs, validating these predictions experimentally across four Transformer architectures.
arXiv:2605. 17877v2 Announce Type: replace Abstract: A significant hurdle for current LLMs is the execution of complex, multi-stage tasks.
arXiv:2605. 15422v3 Announce Type: replace Abstract: Modern RL post-training methods such as GRPO and DAPO train on N response sequences of R tokens sampled from a shared prompt of P tokens, but standard FlashAttention replicates all P prompt tokens N times across both forward and backward passes -- duplicating compute and memory on identical hidden states.