The paper investigates why constrained updates in supervised fine‑tuning (SFT) still lead to forgetting by developing a functional‑path analysis that decomposes changes in old loss into controlled direction, gradient staleness, probe mismatch, and finite‑step effects. It introduces a conditional refresh budget and identifies a restoring direction provided by informative replay, showing that the placement of constraints relative to Adam and the granularity of constraints significantly affect retention across code‑label audits. The study also explores checkpointed generation, replay schedules, and sequential‑task comparisons, demonstrating that subspace regularization, projection, and supervised replay differentially influence backward transfer and content retention.
arXiv:2607. 15587v1 Announce Type: new Abstract: Continual learning studies how deployed language models can continually acquire new tasks without expensive retraining from scratch.
The paper investigates how language models can aid their own continual learning by generating replay data and selecting preservation targets. By organizing task descriptions into Task Banks, the authors show that replay generated by frozen models reduces forgetting on tasks outside the bank and that informed selection of replay targets can improve retention. The study demonstrates that models can support broader self‑preservation, especially when larger models are used, and that replacing new‑task examples with replay can enhance preservation under a fixed example budget.
The paper investigates what information must be preserved in replay buffers for class‑incremental learning. By treating cached predictions as temporally heterogeneous supervision, the authors separate classes known at storage time from those learned later, and evaluate the impact of deleting logit matching. Experiments on CIFAR‑100 with DER++ show that a simple task‑level offset can largely correct the cost of removing later‑class matching, while the cost of disrupting class correspondence remains.
arXiv:2608. 04788v1 Announce Type: cross Abstract: Large language model agents are commonly trained through reinforcement learning with sparse trajectory-level rewards, which offer limited guidance on how strongly individual tokens should be updated.
Large language model agents are commonly trained through reinforcement learning with sparse trajectory-level rewards, which offer limited guidance on how strongly individual tokens should be updated. On-Policy Self-Distillation (OPSD) addresses this by re-scoring generated tokens under a privileged replay view to obtain dense, token-level supervision.
arXiv:2610.06940v1 Announce Type: new Abstract: Continual adaptation of language models can change their output distribution on prompts learned earlier, while retaining every old prompt-answer pair m...
arXiv:2609.21423v1 Announce Type: new Abstract: Online agent deployments produce abundant execution traces, while task-specific verification and expert annotation are costly to scale. We study how to...
arXiv:2609. 04875v1 Announce Type: cross Abstract: Long-running LLM agents are stateful: beyond the transcript they accrete compressed summaries, plaintext memory, pending tool plans, and, under every serving API, a KV cache.
arXiv:2610.34144v1 Announce Type: new Abstract: Fine-tuning adapts pretrained models to new tasks, but identifying which weight changes support the resulting behavior remains difficult. Weight-space...
The paper introduces RECAP, a framework for deciding when to retain, compress, or discard parts of a long-horizon agent’s history to improve future performance. RECAP uses matched context interventions, execution-relational prediction, and independent policy certification to evaluate the value of a failure block relative to the current execution. Experiments across five backbones show that RECAP can edit a significant portion of retained checkpoints while maintaining high continuation success and reducing prompt costs.
arXiv:2605.12922v2 Announce Type: replace Abstract: Large language models can follow complex instructions in a single turn, yet over long multi-turn interactions they often lose the thread of instruc...
CAMR (Compatibility-Aware Memory Replay) is an adaptive framework for continual learning in large language models that jointly manages supervision of new examples and replay of historical data. It uses compatibility checks and a learned memory gate to decide among three actions: general replay of incoming examples, reinforcing compatible related examples, and skipping supervision for mastered duplicates. Evaluated on Qwen3-8B-Base, CAMR outperforms existing methods on WikiBigEdit factual updates and achieves top scores on multiple TRACE tasks while preserving earlier-task performance across diverse domains.