Continual-ARC scores lifelong skill acquisition on ARC-AGI streams
Harness persistence rarely beats a demonstration cache on memorization, and no harness method gains on unseen tasks, unlike weight updates.
Stories per day · 30 days
5 RSS
Relatedlora5agents3fine-tuning3foundation-models3pretraining3reinforcement-learning3benchmarks2catastrophic-forgetting2evaluation2llm-agents2professional-workflows2supervised-fine-tuning2agent-harness1ai-accessibility1arc-agi1benchmark1cifar-1001clif1
50 stories, newest first · one card per story
Stories tagged #continual-learning
TEA-Prompt anchors attention for rehearsal-free continual learning
The method uses 0.38M parameters and stores Gaussian query statistics instead of old images.
LoRA merge audit finds bad adapters inflate forgetting scores
Naive merging of an audited four-adapter set cut GSM8K by 10.5 points and HumanEval by 18.3.
Deletion Tuning matches full fine-tuning by zeroing under 0.05% of weights
A binary mask on gate projections stays near base prior-task accuracy while FFT, LoRA, and random sparse tuning lose up to 5.4 points.
Temporal membership inference attacks on continually updated models
TMIA black-box attack infers whether and when samples were used in training for language and image models
DRAP cuts plasticity in OpenVLA for replay-free continual learning
Risk-weighted gradient memory scales each tensor’s AdamW rate and retains prior LIBERO behaviors better than EWC.
Trajectory raises $45M to make AI model training accessible to nontechnical experts
The startup's continual learning platform lets businesses retrain open-source AI models using user feedback for specialized tasks.
1
BISECT improves low-rank continual learning for Vision Transformers
BISECT partitions rank budget into invariant carrier and novelty complement, outperforming baselines with no extra parameters.
Paper unifies foundation-model pretraining, SFT and RL under one algorithm
A single likelihood trainer mixes static data and RL generations and is reported to improve the accuracy-compute tradeoff.
Surprise-driven continual learning absorbs planted lies about known facts
A model's own prior separates known-fact lies at AUC 0.78–0.83 but not legitimate updates, at AUC 0.53.
Paper unifies pretraining, SFT, and RL under one likelihood algorithm
The authors recast RL stages as maximum-likelihood fits to model-generated targets and mix them in one trainer.
HaMo combines experience cards with continual learning for web agents
The framework beats harness-only and model-only adaptation on within-domain and cross-domain workflow shifts.
Experience replay adapts sEMG gesture models better than EWC
On NinaPro DB1, experience replay reaches 83.3% class-incremental accuracy and limits forgetting to a 1.8-5.8% drop.
Continual Learning Exam shows frontier agents stall on multi-level tasks
The best of 17 setups cleared 47.7% of levels and finished all five levels on 23.3% of tasks.
Continual Learning Exam benchmarks long-horizon learning in LLM agents
Feedback plus one revision lifts median completion by 5.3 points, but no tested frontier setup finishes most five-level tasks.
Concept-consistent continual learning stops saliency explanation drift
C3L anchors regularization to offline class concepts, and authors report gains on five benchmarks against over 20 baselines.
DRIVE weights fine-tuning samples to limit forgetting without source data
Per-sample weights combine Shapley anti-forgetting and adaptation signals with diversity and a prior.
Nestor nested sparsity beats MoE transformers at equal compute
Attention-side sparsity nests with mixture-of-experts sparsity and adapts to new domains with less forgetting.
RCSD raises speculative-decoding acceptance length 35% over HASS
Draft models trained on accepted and rejected tokens report 1.35 times faster inference without a quality loss.
Paper proposes memory-centric continual learning for foundation-model agents
MCCL judges memory by later task gains, with a taxonomy, two-loop lifecycle, eval protocol, and scaling model.
Residual-space adaptation cuts forgetting in continual diffusion personalization
Restricting updates to layers identified by residual dynamics keeps performance while cutting trainable parameters by up to 72%.
WeAct selects LoRA adapters as actions for long-horizon agents
On Qwen3-32B, WeAct beats full fine-tuning by 6.2% and static LoRA merging by 27.5% on eight deep-research benchmarks.
CopilotKit Releases CLiF: A Continual Learning Platform That Turns Agent Traces and User Actions into Skills, Evals, and Fine-Tuning Data
CopilotKit released CLiF, turning agent traces and user actions into skills, evals, and fine-tuning data.
4
Specialists Learn Faster: Knowledge Organization in Pre-training Shapes Knowledge Update Efficiency
Models pretrained on chronological knowledge update faster than shuffled generalists, and a derived law predicts that cost.