The study investigates how gate initialisation and recurrent regularisers (dropout and zoneout) affect learning long dependencies in one‑layer LSTM and GRU networks on synthetic tasks. By running 720 factorial experiments that cross conventional and chrono gate initialisation with up to six regulariser placements, the authors find that chrono initialisation consistently enables networks to reach validation criteria within 5000 updates, whereas conventional initialisation often fails or learns slowly. Dropout generally delays learning, zoneout only helps in limited conditions, and early input sensitivity provides limited predictive power for successful training.
Dropout and zoneout are routinely added to long short-term memory (LSTM) and gated recurrent unit (GRU) networks, but, to our knowledge, their placements have not been crossed with gate initialisation...
arXiv:2609.40063v1 Announce Type: cross Abstract: Low-Rank Adaptive Residual Connections (LARC) give a frozen model a compact numerical state that can learn from feedback. The map $h+BAh$ adds a low-...
The paper investigates how different optimizers acquire distinct generalizing relations in single‑hidden‑layer ReLU networks. By manipulating nuisance‑connected weights and performing a strict‑parity sweep, the authors show that training conditions reshape both shared failures and relative optimizer advantages, with specific interventions improving accuracy on hard stages. The study also compares optimizer behavior across tasks and an external architecture (CIFAR‑10).
arXiv:2610.09835v1 Announce Type: new Abstract: Continual pre-training and fine-tuning in Large Language Models (LLMs) inevitably induce catastrophic forgetting, typically mitigated by replay using o...
arXiv:2601. 18699v2 Announce Type: replace Abstract: Sequential fine-tuning of Large Language Models (LLMs) adaptation to target tasks often triggers catastrophic forgetting, where the acquisition of novel target skills degrades ancestral capabilities.
The study investigates whether a neural network’s training history can better predict its future learning than its current state. Using small multilayer perceptrons trained under 42 different history regimes, the authors measured future learning at four checkpoints via a short probe that replicated the network for 100 updates on a new task. Results show that a low‑dimensional history state does not outperform a calibrated model of the current state for short‑term prediction, but in a companion screen of 1,560 synthetic regression runs, history models forecast final error better than current validation error after 12 epochs and remain comparable until 48 epochs.
The paper investigates when a language model actually learns task information versus when its readout (unembedding matrix) learns to express that information. Using a new parameter‑trajectory cross‑coding method, the authors find that hidden states can encode task data early in training, but the readout reorganizes only in a narrow window (0.4–1.1% of Pythia’s pretraining). By varying the readout learning rate, they show that faster readout learning reduces the gap between hidden‑state encoding and output expression without affecting final performance.
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:2609.37836v1 Announce Type: new Abstract: Neural networks trained toward the same final objective can reach similar predictive performance while retaining internal representations shaped by ear...
arXiv:2506. 21833v2 Announce Type: replace Abstract: Forward-mode automatic differentiation (FmAD) and zero-order (ZO) optimization are increasingly proposed as memory-efficient, backpropagation-free alternatives for large language model (LLM) fine-tuning, yet their benefits are typically evaluated only against standard backpropagation (BP), omitting memory-efficient variants such as activation checkpointing.
The paper introduces LRE (Learned Relevance Eviction), a lightweight, CPU‑only, language‑model‑free scorer that learns which parts of an agent’s interaction history are task‑critical and preserves them verbatim. In experiments, LRE matches or surpasses baseline eviction policies on accuracy‑cost trade‑offs, recovers 93% of full‑history accuracy, reduces worst‑case prompt size by 52%, and outperforms dense and token‑pruning encoders in conversational memory while being 295–1569× smaller. The method also achieves superior budgeted answer quality on LoCoMo reading and can be trained annotation‑free, recovering 95% of supervised scorer performance.
arXiv:2610.16268v2 Announce Type: replace Abstract: Test-time-training (TTT) layers update fast-weight memory while processing an input sequence. A gradient-based update, however, does not by itself...