CL-Bench is the first expert-validated benchmark for continual learning in frontier LLMs across six real-world domains, showing limited gains and that naive in-context learning outperforms dedicated memory systems.
hub Canonical reference
Learning to (Learn at Test Time): RNNs with Expressive Hidden States
Canonical reference. 85% of citing Pith papers cite this work as background.
54
Pith papers citing it
Background
85% of classified citations
abstract
Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, but their performance in long context is limited by the expressive power of their hidden states. We present a practical framework for instantiating sequence modeling layers with linear complexity and expressive hidden states. The key idea is to make the hidden state a machine learning model itself, and the update rule a step of self-supervised learning. Since the hidden state is updated by training even on test sequences, our layers are called Test-Time Training (TTT) layers. We consider two instantiations: TTT-Linear and TTT-MLP, whose hidden state is a linear model and a two-layer MLP respectively. We evaluate our instantiations at the scale of 125M to 1.3B parameters, comparing with a strong Transformer and Mamba, a modern RNN. Similar to Transformer, TTT-Linear and TTT-MLP can keep reducing perplexity by conditioning on more tokens, while Mamba cannot after 16k context. TTT-MLP still faces challenges in memory I/O, but shows larger potential in long context, pointing to a promising direction for future research.
hub tools
citation-role summary
background 12
method 1
citation-polarity summary
representative citing papers
WriteSAE introduces sparse autoencoders with rank-1 matrix atoms for recurrent state updates, allowing replacement tests that outperform deletion on 92.4% of positions and a formula predicting logit changes with R²=0.98.
Tapered Language Models monotonically decrease MLP width across depth with a cosine schedule, yielding better perplexity and downstream performance than uniform-width baselines across multiple architectures and scales at no extra cost.
AURA-Mem uses an action-gated recurrent memory trained on closed-loop action error to deliver constant 4,224-byte state and 5-9x fewer writes than baselines while matching base policy success on LIBERO-Long.
TTRL gains are reinterpreted as mostly sharpening rather than learning, with an identified extinction window causing net corruption; TTRL-Guard mitigates via FRS, MPS, and RCSU for improved pass@1.
Mamba-VGGT introduces a Sliding Window Mamba memory module and Zero-Init Spatial Memory Injector to enable persistent long-range geometric reasoning in VGGT for extended video sequences.
Mem3R achieves better long-sequence 3D reconstruction by decoupling tracking and mapping with a hybrid memory of TTT-updated MLP and explicit tokens, reducing model size and trajectory errors.
GC-TTT adapts goal-conditioned policies at test time by fine-tuning on self-supervised selected goal-related offline data, yielding performance gains in loco-navigation and manipulation tasks.
SDM sparsifies the Gated DeltaNet update rule to enable 1000x larger recurrent memory states at iso-FLOP, improving long-context recall and short-context reasoning over GDN and matching full attention at 8B scale.
Anchoring Gaussian centers to predicted raymaps and jointly optimizing RGB, raymap, and camera losses with a dual-frequency curriculum suppresses pose drift and improves pose-free 3D reconstruction on long sequences.
FFN performs efficient test-time training on multi-hour videos by forgetting the exiting frame, anticipating the next, and adapting only when a surprise metric exceeds a dynamic threshold.
RoboMME-Interference measures VLA memory under cross-session interference and finds that all tested systems decay to near their no-memory baseline as unrelated sessions accumulate.
ARTS improves automated scientific discovery by using reasoning LMs with test-time training to separate hypothesis merit from execution quality in tree search, achieving 15.3% relative gains on 22 MLGym and MLEBench tasks.
TTT-NTP adapts pretrained LLMs at test time by training fast weights to match next-position hidden states from the forward pass, yielding consistent gains on long-context benchmarks across Llama, Mistral, and Qwen models.
Reinforcement learning after SFT conversion narrows the performance gap between sliding-window attention and full self-attention on math reasoning benchmarks while preserving linear complexity.
U-TTT adds spatial and frequency-domain test-time training layers to a U-Net architecture to adapt PET denoising models at inference time for improved generalization across unseen dose levels and scanners.
Sleep-time Knowledge Seeding plus Dreaming lets LLMs expand capacity, distill fragile in-context memories into stable parameters, and self-improve without human labels.
Introduces AgentOdyssey, a procedural generator of open-ended long-horizon text games, to evaluate test-time continual learning agents and diagnose limits in exploration, memory, and planning.
Proves SLiCEs are universal time-series generators approximating path laws in W_∞ and proposes G-SLiCEs for path-space flow matching with benefits on irregular grids.
EvoLib improves black-box LLM test-time performance by maintaining an evolving, self-scored library of reusable skills and insights, without parameter updates or ground-truth feedback.
OSDN adds online diagonal preconditioning to the Delta Rule, preserving chunkwise parallelism while proving super-geometric convergence and delivering 32-39% recall gains at 340M-1.3B scales.
A 130M-parameter 1-layer GPN achieves FineWeb-Edu perplexity 18.06, within 13% of a 12-layer Transformer++ (16.05) and 18% of a 10-layer GDN (15.34).
Converts pretrained Vision Transformers to linear-complexity TTT models via architectural and representational alignment, demonstrated by linearizing Stable Diffusion 3.5 with 1-hour fine-tuning to match quality at 1.32-1.47x faster inference.
LLM agents trained with a task-success reward on self-generated knowledge can spontaneously explore and adapt to new environments without any rewards or instructions at inference, yielding 20% gains on web tasks and allowing a 14B model to beat Gemini-2.5-Flash.
citing papers explorer
-
Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
CL-Bench is the first expert-validated benchmark for continual learning in frontier LLMs across six real-world domains, showing limited gains and that naive in-context learning outperforms dedicated memory systems.
-
WriteSAE: Sparse Autoencoders for Recurrent State
WriteSAE introduces sparse autoencoders with rank-1 matrix atoms for recurrent state updates, allowing replacement tests that outperform deletion on 92.4% of positions and a formula predicting logit changes with R²=0.98.
-
Tapered Language Models
Tapered Language Models monotonically decrease MLP width across depth with a cosine schedule, yielding better perplexity and downstream performance than uniform-width baselines across multiple architectures and scales at no extra cost.
-
AURA: Action-Gated Memory for Robot Policies at Constant VRAM
AURA-Mem uses an action-gated recurrent memory trained on closed-loop action error to deliver constant 4,224-byte state and 5-9x fewer writes than baselines while matching base policy success on LIBERO-Long.
-
Detecting and Mitigating the Correct-Answer Extinction Window in Test-Time Reinforcement Learning with Majority Voting
TTRL gains are reinterpreted as mostly sharpening rather than learning, with an identified extinction window causing net corruption; TTRL-Guard mitigates via FRS, MPS, and RCSU for improved pass@1.
-
Mamba-VGGT: Persistent Long-Sequence Video Geometry Grounded Transformer via External Sliding Window Mamba Memory
Mamba-VGGT introduces a Sliding Window Mamba memory module and Zero-Init Spatial Memory Injector to enable persistent long-range geometric reasoning in VGGT for extended video sequences.
-
Mem3R: Streaming 3D Reconstruction with Hybrid Memory via Test-Time Training
Mem3R achieves better long-sequence 3D reconstruction by decoupling tracking and mapping with a hybrid memory of TTT-updated MLP and explicit tokens, reducing model size and trajectory errors.
-
Test-time Offline Reinforcement Learning on Goal-related Experience
GC-TTT adapts goal-conditioned policies at test time by fine-tuning on self-supervised selected goal-related offline data, yielding performance gains in loco-navigation and manipulation tasks.
-
Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
SDM sparsifies the Gated DeltaNet update rule to enable 1000x larger recurrent memory states at iso-FLOP, improving long-context recall and short-context reasoning over GDN and matching full attention at 8B scale.
-
NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction
Anchoring Gaussian centers to predicted raymaps and jointly optimizing RGB, raymap, and camera losses with a dual-frequency curriculum suppresses pose drift and improves pose-free 3D reconstruction on long sequences.
-
Forget, Anticipate and Adapt: Test Time Training for Long Videos
FFN performs efficient test-time training on multi-hour videos by forgetting the exiting frame, anticipating the next, and adapting only when a surprise metric exceeds a dynamic threshold.
-
RoboMME-Interference: Benchmarking Robot Memory Under Interference
RoboMME-Interference measures VLA memory under cross-session interference and finds that all tested systems decay to near their no-memory baseline as unrelated sessions accumulate.
-
Learning the ARTS of Search for Automated Discovery
ARTS improves automated scientific discovery by using reasoning LMs with test-time training to separate hypothesis merit from execution quality in tree search, achieving 15.3% relative gains on 22 MLGym and MLEBench tasks.
-
Test-Time Training with Next-Token Prediction
TTT-NTP adapts pretrained LLMs at test time by training fast weights to match next-position hidden states from the forward pass, yielding consistent gains on long-context benchmarks across Llama, Mistral, and Qwen models.
-
Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning
Reinforcement learning after SFT conversion narrows the performance gap between sliding-window attention and full self-attention on math reasoning benchmarks while preserving linear complexity.
-
U-TTT: Towards Generalizable PET Image Denoising via Test-Time Training
U-TTT adds spatial and frequency-domain test-time training layers to a U-Net architecture to adapt PET denoising models at inference time for improved generalization across unseen dose levels and scanners.
-
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories
Sleep-time Knowledge Seeding plus Dreaming lets LLMs expand capacity, distill fragile in-context memories into stable parameters, and self-improve without human labels.
-
AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents
Introduces AgentOdyssey, a procedural generator of open-ended long-horizon text games, to evaluate test-time continual learning agents and diagnose limits in exploration, memory, and planning.
-
Universal Time Series Generation with Neural Controlled Differential Equations
Proves SLiCEs are universal time-series generators approximating path laws in W_∞ and proposes G-SLiCEs for path-space flow matching with benefits on irregular grids.
-
Test-Time Learning with an Evolving Library
EvoLib improves black-box LLM test-time performance by maintaining an evolving, self-scored library of reusable skills and insights, without parameter updates or ground-truth feedback.
-
OSDN: Improving Delta Rule with Provable Online Preconditioning in Linear Attention
OSDN adds online diagonal preconditioning to the Delta Rule, preserving chunkwise parallelism while proving super-geometric convergence and delivering 32-39% recall gains at 340M-1.3B scales.
-
A Single-Layer Model Can Do Language Modeling
A 130M-parameter 1-layer GPN achieves FineWeb-Edu perplexity 18.06, within 13% of a 12-layer Transformer++ (16.05) and 18% of a 10-layer GDN (15.34).
-
Linearizing Vision Transformer with Test-Time Training
Converts pretrained Vision Transformers to linear-complexity TTT models via architectural and representational alignment, demonstrated by linearizing Stable Diffusion 3.5 with 1-hour fine-tuning to match quality at 1.32-1.47x faster inference.
-
Training LLM Agents for Spontaneous, Reward-Free Self-Evolution via World Knowledge Exploration
LLM agents trained with a task-success reward on self-generated knowledge can spontaneously explore and adapt to new environments without any rewards or instructions at inference, yielding 20% gains on web tasks and allowing a 14B model to beat Gemini-2.5-Flash.
-
DexWorldModel: Causal Latent World Modeling towards Automated Learning of Embodied Tasks
CLWM with DINOv3 targets, O(1) TTT memory, SAI latency masking, and EmbodiChain training achieves SOTA dual-arm simulation performance and zero-shot sim-to-real transfer that beats real-data finetuned baselines.
-
Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction
Scal3R achieves better accuracy and consistency in large-scale 3D scene reconstruction by maintaining a compressed global context through test-time adaptation of lightweight neural networks on long video sequences.
-
In-Place Test-Time Training
In-Place TTT adapts LLM MLP projection matrices at test time with a next-token-aligned objective and chunk-wise updates, enabling better long-context performance as a drop-in enhancement.
-
Listen to Rhythm, Choose Movements: Autoregressive Multimodal Dance Generation via Diffusion and Mamba with Decoupled Dance Dataset
LRCM is a new multimodal diffusion model with audio and text Conformers plus Motion Temporal Mamba for generating long, coherent dance sequences from rhythm and descriptions using a decoupled dataset.
-
Higher-order Linear Attention
Higher-order Linear Attention realizes second-order and higher interactions in linear-time causal attention via constant-size state and associative scans.
-
Kimi Linear: An Expressive, Efficient Attention Architecture
Kimi Linear hybridizes linear attention with a new KDA module to beat full attention on tasks while slashing KV cache by 75% and speeding decoding up to 6x.
-
Short window attention enables long-term memorization
Short sliding windows in hybrid attention-xLSTM models boost long-context performance by encouraging long-term memory use, and stochastic window sizing improves both short and long tasks.
-
ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution
ShinkaEvolve improves sample efficiency in LLM-driven program evolution via parent sampling, code novelty rejection-sampling, and bandit LLM ensemble selection, achieving new SOTA circle packing with 150 samples and gains on math reasoning and competitive programming tasks.
-
MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
MiniMax-M1 is a 456B parameter hybrid-attention MoE model trained with CISPO RL that achieves performance comparable or superior to DeepSeek-R1 and Qwen3-235B on reasoning and software engineering tasks while training in three weeks on 512 GPUs.
-
Test-Time Training Done Right
Large-chunk online updates during inference let test-time training scale state capacity to 40% of model size and handle contexts up to 1M tokens without custom kernels.
-
Titans: Learning to Memorize at Test Time
Titans combine attention for current context with a learnable neural memory for long-term history, achieving better performance and scaling to over 2M-token contexts on language, reasoning, genomics, and time-series tasks.
-
Sakana Fugu Technical Report
Sakana Fugu trains LLM orchestrators using fine-tuning, evolutionary algorithms, and RL to build query-adaptive multi-agent scaffolds, claiming SOTA results on benchmarks including SWE-Bench Pro and GPQA-Diamond.
-
Q-Delta: Beyond Key-Value Associative State Evolution
Q-Delta extends linear attention by introducing a query-conditioned delta rule that incorporates mixed key-query errors into recurrent state updates for improved stability and performance.
-
Latent Recurrent Transformer: Architecture Exploration, Training Strategies, and Scaling Behavior
Latent Recurrent Transformer augments autoregressive transformers with a cross-layer recurrent latent pathway from prior hidden states and uses interleaved parallel training to improve loss and in-context learning at ~0.3% extra parameters.
-
SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer
SANA-WM is a 2.6B-parameter efficient world model that synthesizes minute-scale 720p videos with 6-DoF camera control, trained on 213K public clips in 15 days on 64 H100s and runnable on single GPUs at 36x higher throughput than prior open baselines.
-
Mela: Test-Time Memory Consolidation based on Transformation Hypothesis
Mela is a Transformer variant with a dual-frequency Hierarchical Memory Module and MemStack that performs test-time memory consolidation, outperforming baselines on long contexts.
-
Cortico-cerebellar modularity as an architectural inductive bias for efficient temporal learning
CB-RNNs with a cerebellar feedforward module learn temporal tasks faster than matched RNNs, with the module driving efficiency even after freezing the recurrent core as a fixed reservoir.
-
Kaczmarz Linear Attention
Kaczmarz Linear Attention replaces the empirical coefficient in Gated DeltaNet with a key-norm-normalized step size derived from the online regression objective, yielding lower perplexity and better needle-in-haystack performance.
-
Federated Nested Learning: Collaborative Training of Self-Referential Memories for Test-Time Adaptation
FedNL reformulates federated learning as nested optimization with linear attention for collaborative test-time adaptation on non-IID data.
-
PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents
PACEvolve++ uses a phase-adaptive reinforcement learning advisor to decouple hypothesis selection from execution in LLM-driven evolutionary search, delivering faster convergence than prior frameworks on load balancing, recommendation, and protein tasks.
-
Measuring Accuracy and Energy-to-Solution of Quantum Fine-Tuning of Foundational AI Models
Trapped-ion quantum fine-tuning of AI models shows linear energy scaling and 24% better classification error than classical logistic regression or SVM baselines, with a projected energy break-even at 34 qubits.
-
Understand and Accelerate Memory Processing Pipeline for Large Language Model Inference
Unifying LLM long-context optimizations into a four-step memory pipeline and offloading sparse/irregular stages to FPGA yields up to 2.2× speedup and 4.7× energy savings versus GPU-only.
-
Attention Residuals
Attention Residuals replaces fixed residual summation with input-dependent softmax attention over preceding layers, and a blocked variant is shown to improve uniformity and downstream performance in a 48B-parameter model pre-trained on 1.4T tokens.
-
Long-LRM++: Preserving Fine Details in Feed-Forward Wide-Coverage Reconstruction
Long-LRM++ achieves real-time 14 FPS high-fidelity 360-degree scene reconstruction from 32-64 views by using semi-explicit Gaussians plus a light decoder, matching LaCT quality on DL3DV and improving depth prediction.
-
Nirvana: A Specialized Generalist Model With Task-Aware Memory Mechanism
Nirvana adds a task-aware memory trigger and updater to specialized generalist models, achieving strong general benchmark results, lowest perplexity in biomedicine/finance/law, and improved MRI reconstruction fidelity.
-
TTT3R: 3D Reconstruction as Test-Time Training
TTT3R derives a closed-form learning rate from memory-observation alignment confidence to boost length generalization in RNN-based 3D reconstruction by 2x in global pose estimation.