Stable Recovery Manifold hypothesis: catastrophic forgetting as accessibility problem, not information destruction
A new arXiv preprint investigates the geometric structure of recoverability in continual learning using Split CIFAR-100 and a sequentially trained ResNet-18. The authors introduce Recovery Subspace Dimensionality (k_t) and find that recovery dimensionality remains stable across tasks (mean k_t = 8.0) despite substantial representational drift, with principal-angle drift strongly predicting recoverability (r = -0.862). The findings support the Stable Recovery Manifold hypothesis: forgotten knowledge remains compactly decodable, reframing catastrophic forgetting as a manifold-alignment and accessibility problem rather than true information loss.
Related guides (1)
Related events (8)
Self-Generated Replay Nearly Eliminates Catastrophic Forgetting in Language Models
This paper investigates catastrophic forgetting in language models during continual learning, finding that models can use self-generated samples from their own training distribution as effective replay data, nearly eliminating forgetting without requiring stored exemplars. The authors identify two key conditions where forgetting persists: when models are pretrained near capacity saturation (leaving no room for new knowledge), and when low learning rates are used to reduce forgetting at the cost of requiring far more training steps. Self-generated replay breaks this learning-rate/forgetting tradeoff, enabling fast high-learning-rate finetuning without degradation on prior tasks.
QK-Restore: Fixing long-context recall degradation caused by CoT fine-tuning in hybrid LLMs
Researchers find that chain-of-thought supervised fine-tuning systematically degrades long-context recall in hybrid linear-attention models (HypeNet, Jet-Nemotron), with Needle-In-A-Haystack performance collapsing dramatically—e.g., HypeNet-9B dropping from 67.2% to 9.4% at 256K context. The root cause is identified as CoT-SFT biasing attention gradients toward short-range patterns, corrupting the query-key projections responsible for long-range routing. The paper proposes QK-Restore, a training-free fix that restores only W_Q and W_K from the pre-SFT checkpoint, recovering long-context capability while preserving reasoning gains.
SETA: Sparse Subspace-to-Expert Sharing for Continual Learning in LLMs
Researchers introduce SETA (Mixture of Sparse Experts for Task Agnostic Continual Learning), a framework addressing catastrophic forgetting in LLMs via adaptive sparse subspace decomposition into task-specific and shared expert modules. The approach uses adaptive elastic anchoring and routing-aware regularization to protect shared knowledge at both weight and routing levels. Experiments on LLaMA-2 7B and Qwen3-4B show competitive or superior performance versus continual learning baselines, with strong retention of early-task knowledge.
Study finds SAE unstable features reflect reproducible subspaces, not pure noise
A new arXiv paper investigates feature stability in sparse autoencoders (SAEs), measuring the probability that individual learned features reappear across independent training runs. The authors find a functional asymmetry: stable features carry most reconstruction-relevant signal, while unstable features are individually non-reproducible but concentrate in reproducible lower-rank subspaces, suggesting seed dependence reflects basis ambiguity rather than noise. A synthetic model confirms that low-rank ground-truth features can be recovered at the subspace level even when individual SAE latents are non-identifiable across seeds. The work has direct implications for interpretability research that relies on SAE features as meaningful, stable units of analysis.
AdamO optimizer and dynamical isometry regularization preserve plasticity in continual learning
A new arXiv preprint connects plasticity loss in continual learning to the empirical Neural Tangent Kernel and identifies dynamical isometry—keeping layer-wise Jacobian singular values near one—as a key mechanism for maintaining learning capacity under non-stationarity. The authors propose an isometry-promoting regularization scheme that can reactivate dormant ReLU units and introduce AdamO, an Adam-style optimizer that decouples isometry regularization from gradient updates analogously to AdamW. The methods are evaluated on supervised and reinforcement-learning continual-learning benchmarks, consistently matching or outperforming prior approaches. The work also reinterprets existing plasticity-preserving methods as targeting only partial isometry measures.
Beyond Prediction Accuracy: Target-Space Recovery Profiles for Evaluating Model-Brain Alignment
This paper introduces a framework for evaluating alignment between artificial vision models and the human visual cortex that goes beyond scalar prediction accuracy. Using repeated fMRI data from the Natural Scenes Dataset, the authors decompose brain response spaces into reproducible dimensions and measure which of these dimensions are recovered by model predictions. A key finding is that pretrained and randomly initialized models can achieve similar prediction accuracy while showing distinct recovery profiles, revealing that accuracy alone can mask fundamental model-brain mismatches. The framework also enables brain-to-brain comparisons as a diagnostic human reference baseline.
Latent World Recovery: multimodal learning framework for missing modalities in bioscience
A new arXiv preprint introduces Latent World Recovery (LWR), a framework for multimodal learning when some modalities are unavailable at training or inference time. LWR aligns modality-specific embeddings in a shared latent space and fuses only available modalities, avoiding explicit reconstruction of missing ones. The approach is evaluated on incomplete multi-omics benchmarks for cancer phenotype classification and survival prediction, demonstrating robustness under partial observation.
Failed reasoning traces encode recoverability structure for test-time routing and post-training analysis
A new arXiv paper argues that failed reasoning traces from post-trained LLMs contain exploitable signal about whether failures are recoverable via resampling or require structural intervention. The authors derive three trajectory features from the distributional signature of failed rollouts (not their text content) that cluster failures into stable regimes and characterize failure topography across post-training methods with 84.3% accuracy. A training-free routing rule built on these features lifts rescue rates by +12.2% on a deployment-relevant hard subset, and the features transfer across model families. The work reframes failed traces as diagnostic objects rather than discarded data, with implications for inference-time compute allocation and post-training analysis.