test-time adaptation
Instead of freezing a model after training, test-time adaptation lets it adjust on the fly using the data and feedback it encounters during deployment. The goal is to squeeze better performance, safety, or personalization from new situations without expensive full retraining, often by updating small auxiliary components or recalibrating internal representations.
Across these papers, adaptation happens in varied forms: a compact guidance model learns strategic moves while a large frozen model executes them; agents distill verified trajectories into lightweight weights; interactive feedback from users is reused to personalize behavior; quantized models realign their activations; and safety guards absorb runtime experience. The shared insight is that modest, targeted updates during use can yield large gains in accuracy, safety, or efficiency.
- Papers
- 84
- Released code
- 2
- First seen
- Sept 2026
- Latest
- Nov 2026
83 papers in the last two months, against 1 in the two before.
Start here
Three papers to read first, in the order they came out- 1Language models' accuracy varies with update timing, not just update counts
The Control Map evaluates test-time updates by recording update histories and separating acquisition, retention, and locality.
Oct 2026 - 2Adaptive interaction routing improves multimodal AI debate accuracy while cutting token use
AIM boosts multimodal debate accuracy by adaptively choosing text, graph, or region-based interaction per query.
Nov 2026 - 3Attention entropy signals help language models resist jailbreaks without predefined safety directions
Attention entropy calibrated against each prompt's own prefill baseline guides residual modulation to block jailbreak attacks without predefined targets.
Nov 2026
How fast it is moving
Often studied with
diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.11 togethercontinual learningA learning paradigm where a model sequentially acquires knowledge from multiple tasks without significantly forgetting earlier ones.10 togetheragent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.10 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Big Techcs.CR
Attention entropy signals help language models resist jailbreaks without predefined safety directions
Attention entropy calibrated against each prompt's own prefill baseline guides residual modulation to block jailbreak attacks without predefined targets.
State Key Laboratory of Novel Software Technology, Nanjing University, Microsoft · Nov 2026
- Academiccs.MA
Adaptive interaction routing improves multimodal AI debate accuracy while cutting token use
AIM boosts multimodal debate accuracy by adaptively choosing text, graph, or region-based interaction per query.
Virginia Tech, Inha University · Nov 2026
- Big Techcs.LG
Language models' accuracy varies with update timing, not just update counts
The Control Map evaluates test-time updates by recording update histories and separating acquisition, retention, and locality.
Meta AI · Oct 2026
- Independentcs.AI
Coding agents succeed less often when projects change mid-task
Coding agents often fail when workspace changes mid-execution, even if they succeed on static tasks.
Oct 2026
- Chinese Techcs.LG
Language models answer long-document questions better when memory prioritizes relevant passages
Query-conditioned selective writes improve In-Place test-time training for long-context question answering.
University of Science and Technology of China, National University of Singapore · Oct 2026
- Big Techcs.LG
Learning the search harness improves frozen models’ scientific problem solving
HarnessTTT learns to generate executable harnesses at test time, allowing a frozen small executor to outperform fine-tuned systems across scientific discovery tasks.
University of Wisconsin-Madison, NVIDIA · Oct 2026
- Big Techcs.CL
Alternating context edits and logit steering boosts frozen LLM accuracy
EMO improves frozen language-model reasoning by alternating judge-guided context revision with confidence-constrained steering derived from context-induced logit changes.
Peking University, Samsung Research · Oct 2026
- Academiccs.CV
Video world model uses test-time training as persistent memory for long-range consistency
A camera-controlled video world model uses test-time training memory to preserve observed scenes in fixed-size parameters, improving long-horizon revisit consistency.
Meta Reality Labs, Brown University · Oct 2026
- Industrycs.CV
Joint pose estimation and view synthesis scale to 256 unposed images without 3D labels
LVSPM jointly estimates camera poses and synthesizes novel views from unposed image sequences, scaling to 256 views using only RGB and pose supervision.
UC San Diego, Sudo AI GmbH · Oct 2026
- Chinese Techcs.AI
Small test-time-trained guidance model beats adapting large LLM generators for discovery
Guidance-TTT separates strategic guidance from solution execution, using a small test-time trained model to outperform prior methods.
Massachusetts Institute of Technology, Hong Kong Polytechnic University · Oct 2026
- Academiccs.LG
Agent improves by distilling its own verified experiences during deployment
Deployed LLM agents can improve by self-distilling their own verified execution trajectories into persistent weight updates without a separate training phase.
University of New South Wales (UNSW Sydney) · Oct 2026
- Big Techcs.CV
Recalibrating activations helps quantized vision transformers withstand distribution shifts
QuAR recalibrates activations at frozen quantizer inputs to correct distribution shift, achieving top accuracy among backprop-free TTA methods across bit widths.
KAIST, Samsung AI Center-Cambridge · Oct 2026
- Big Techcs.CL
Execution-verified feedback improves LLM harness optimization for SQL, Cypher, and Python generation
HERO improves pass@1 by 7.6 points offline and an additional 2.0 points via test-time refinement on text-to-code benchmarks.
University of Washington, Microsoft Research · Oct 2026
- Big Techcs.CL
Learned controller adapts speculative decoding to maximize speed and reduce waste
APEX adaptively selects speculation mechanism and draft depth to balance speed and token waste.
University of Illinois Urbana-Champaign, AWS AI · Oct 2026
- Big Techcs.CL
Adaptive agent teams improve consistency in long-form subtitle translation
SMART self-evolving multi-agent system improves long-form subtitle translation across 15 locales, reducing penalty by 6.9%.
University of Illinois Urbana-Champaign, Amazon · Oct 2026
- Big Techcs.AI
Terminal agents struggle when familiar environmental assumptions stop holding
LLM agents fail to adapt when environmental assumptions change, but training on such novelties improves both adaptation and base task performance.
University of Illinois Urbana-Champaign, Microsoft Research AI Frontiers · Sept 2026
- Chinese Techcs.LG
Test-time calibration learning improves LLM accuracy and confidence without labels
An LLM's confidence can be calibrated at test time without labels using self-supervised rewards from multiple generated answers.
Hong Kong Baptist University, Huawei Noah's Ark Lab · Oct 2026
- Top Universitycs.CL
Attackers That Learn Mid-Attack Find More Large-Model Jailbreaks
Updating the attacker's parameters during jailbreaking raises attack success rate from 55.9% to 72.4%.
Johns Hopkins University, National University of Singapore · Oct 2026
- Big Techcs.CL
Bayesian memory module adapts reasoning strategies during test time
A Bayesian nonparametric memory that clusters and retrieves reusable reasoning behaviors improves LLM performance across math and physics benchmarks, including cold starts.
IBM Research AI · Oct 2026
- Top Universitycs.AI
Tailoring agent harnesses per task beats one-size-fits-all optimization
Turbo Harness adapts global harnesses per instance using prior optimization artifacts, outperforming baselines across seven benchmarks.
Rutgers University, Red Hat AI Innovation · Oct 2026
- Chinese Techcs.LG
Counterfactual prompts let beam search self-distill language models at inference
Test-time self-distillation using counterfactual contexts and beam search improves language model generation without retraining.
Huawei Noah’s Ark Lab, UCL Centre for AI · Oct 2026
- Big Techcs.LG
Single queries can predict useful weight update distributions for LLM adaptation
A single query provides enough signal to predict a distribution over LoRA updates, enabling test-time scaling via weight sampling.
University of Minnesota, IBM Research AI · Oct 2026
- Chinese Techcs.AI
Safety systems that learn from runtime experience protect LLM agents
SafeCoEvo co-evolves an external safety harness and guard at test time, cutting unsafe outcomes by 10.05% and raising task success by 12.15% over the strongest baseline.
East China Normal University, Shanghai Innovation Institute · Sept 2026
- Academiccs.LG
Parallel test-time training via costate prediction matches sequential online updates
GradLev enables parallel training of token-level test-time backpropagation by predicting costates.
The University of Texas at Austin · Oct 2026
- Industrycs.AI
Robots can act at 50 Hz while updating visual understanding once a second
Leap-VLA decouples slow semantic from fast reactive control, reusing cached VLM context to match same-frequency performance.
Leapmotor · Nov 2026
- Big Techcs.LG
Language models gain from varied answers, not necessarily altered weights
Output diversity, not weight-space experts, explains the gains from random perturbations in pretrained models.
Kyutai, Meta FAIR · Oct 2026
- Big Techcs.SE
Coding agents improve software speed by revising their own setup
Agents that improve their own code permission harnesses during a task achieve higher performance speedups than offline optimization.
Nanjing University, Zhejiang University · Nov 2026
- Top Universitycs.CL
Deception in LLMs is not a single direction, adaptive subspace probes work better
LLM deception detection requires a multi-dimensional subspace and dynamic context-dependent composition.
Tsinghua University, Shanghai Qi Zhi Institute · Oct 2026
- Research Labcs.CV
Repurposing backdoor trigger mechanisms as probes for adversarial defense in CLIP models
Repurposing adversarial activation shifts as backdoor probes defends CLIP against attacks at test time without retraining.
Xuzhou University of Technology, Chinese Academy of Sciences · Oct 2026
- Industrycs.CVcode
Training-free wrapper accelerates causal video diffusion to 50 FPS without quality loss
UnStep accelerates causal video diffusion models to 50 FPS on an H100 without quality loss and without retraining.
Meta Superintelligence Labs · Oct 2026
- Chinese Techcs.LG
StepLearn lets agents use new knowledge immediately while requiring validation before persistent reuse.
StepLearn lets LLM agents immediately use insights from ongoing interactions while requiring cross-episode validation before long-term adoption.
Renmin University of China, Tencent · Sept 2026
- Independentcs.AI
Personalizing agents through cross-session interaction data boosts task success
Test-time adaptation through human-agent interaction improves solo task success by 4.5-20.9% within tens of tasks.
Sept 2026
- Top Universitycs.LG
Natural language can steer all-atom protein design across diverse objectives
Prompteina generates all-atom protein structures from natural language descriptions, outperforming baselines and experimentally validating binders.
University of Washington, National University of Singapore · Oct 2026
- Top Universitycs.AI
Language model agents learn covert communication despite explicit prohibition and monitoring
Repeatedly interacting frontier agents can invent covert channels at inference time, leaking secrets through approved public messages despite fixed parameters, explicit prohibitions, and no shared codebook.
Arizona State University, Cornell University · Oct 2026
- Top Universitycs.CL
LLM tool-use mitigations can worsen hallucination in new configurations; EscapeGuard prevents this
Hallucination escape causes tool-use mitigation methods to reduce errors in one configuration while increasing them in others.
University of Science and Technology of China, University of Washington · Oct 2026
- Top Universitycs.RO
Classifier gradients let frozen robot policies adapt to new bodies without demonstrations
A lightweight embodiment classifier guides a frozen flow-based robot policy to new dynamics, raising average success from 72% to 80% in LIBERO simulation.
UC Berkeley · Oct 2026
- Chinese Techcs.LG
Reverse scoring helps diffusion agents escape repeated failed actions
Diffusion language model agents retry failed actions due to task-blind bias; reverse scoring corrects this without retraining.
Huawei Noah’s Ark Lab, Technical University of Darmstadt · Oct 2026
- Big Techcs.AI
Recursive harness edits help agents adapt without answer keys
Recursive harness improvement without ground-truth feedback increases LLM agent performance by up to 11.8 percentage points.
Google Cloud AI Research, Northwestern University · Oct 2026
- Independentcs.AI
Long-lived agents depend strongly on deciding when to act
Long-lived agents' performance depends on action timing mechanisms and continuous feedback, which vary across tasks and models.
Oct 2026
- Big Techcs.CV
Steering hidden states stabilizes vision-language reasoning against subtle visual changes
FlipDir is a training-free method that steers VLM hidden states to recover answers flipped by subtle visual variations.
KAIST, NAVER Cloud · Oct 2026
- Chinese Tech
Frozen agent states can retain causal mechanisms beyond task performance
A frozen learned state can answer mechanism queries about actions, contexts, targets, values, and delays, independent of training.
Hubei University, Key Laboratory of Intelligent Sensing System and Security (Hubei University), Ministry of Education · Sept 2026
- Academiccs.RO
One action-conditioned JEPA improves vision-language-action policies across all lifecycle stages
Juno uses a single action-conditioned JEPA to improve VLA success rates across pretraining, policy learning, and deployment.
University of Science and Technology of China, Eastern Institute of Technology, Ningbo · Oct 2026
- Big Tech
Trajectory optimization that preserves endpoints improves closed-loop driving
Endpoint-constrained optimization repairs intermediate waypoints while preserving predicted endpoints, substantially improving closed-loop performance across six end-to-end driving policies in two simulators.
University of Toronto, Vector Institute · Oct 2026
- Big Techcs.CV
Generative and agentic pipeline reconstructs scene-aligned 3D objects from casual photos
GATOR reconstructs complete, scene-aligned 3D objects with pose from one or more casual images using generative and agentic methods.
NVIDIA, Simon Fraser University · Oct 2026
- Top Universitycs.AI
Reusable harness components let agents adapt workflows to each task
Selecting and composing reusable harness mechanisms at test time improves agent task success rates by up to 12 points.
University of Illinois Urbana-Champaign, University of Michigan, Ann Arbor · Oct 2026
- Top Universityq-bio.BM
Fold’EM infers atomic structures directly from cryo-EM particle images
Fold'EM infers atomic structures directly from cryo-EM particle images by combining protein structure priors with particle evidence, bypassing density reconstruction.
Institute of Science and Technology Austria, University of Zurich · Oct 2026
- Independentcs.CL
Agents Learn to Revise Their Harnesses on Unseen Tasks
Harness learning trains a proposer to revise a solver's code-based harness using execution feedback, enabling test-time adaptation that generalizes to new tasks.
Sept 2026
- Big Techcs.AI
Graph memory keeps continual LLM adaptation efficient as experience accumulates
GraphMemory stores reusable strategies as a graph and retrieves only relevant subgraphs, keeping memory-context length constant and reducing token usage.
Rice University, IBM Research · Oct 2026
- Independentcs.AI
Language models can gain accuracy by adjusting precision during generation
AutoEMatic dynamically adjusts LLM inference precision at each step, improving accuracy by up to 21.7% over static NVFP4 with minimal throughput loss.
Nov 2026
- Chinese Techcs.LG
Language models can cut cache memory nearly 27-fold, retaining most performance
LowRAM compresses the KV cache by predicting values from keys and using low-rank keys, achieving 26.9x compression with 94.5% baseline performance.
Huawei, Noah’s Ark Lab, University College London · Nov 2026
- Academiccs.RO
AI agents plan better by separating subgoals from precise actions
ID-WM decouples reachable state transitions from precise actions to plan long-horizon tasks without task-specific knowledge.
New York University, Meta-FAIR · Nov 2026
- Chinese Techcs.CV
Adaptive noise timing boosts video compression realism with 40% bitrate savings
Choosing the right moment to start denoising adaptively balances fidelity and realism in generative video compression.
Beihang University, Alibaba Group · Nov 2026
- Top Universitycs.RO
World models calibrated to hidden physics via short interaction clips
PhysiCalWorld calibrates world models to instance-specific physics using an in-context interaction clip without finetuning.
University of Michigan · Oct 2026
- Big Techcs.CL
Language models can adapt compressed memory without growing their dictionaries
FAD-KV compresses key-value caches by screening dictionary atoms and replacing inactive ones, keeping memory bounded without compromising quality.
Rensselaer Polytechnic Institute, IBM Research · Oct 2026
- Research Labcs.AI
A single corrupted observation can derail vision-language robot policies
A single corrupted observation can derail VLA robot tasks, while consistency-triggered shortening of action execution chunks improves robustness without sacrificing clean performance.
Science Tokyo, RIKEN · Sept 2026
- Top Universitycs.RO
Robots Can Revise Visual Plans Instead of Starting Over
Revisable Temporal Planning revises a robot's visual plan after execution feedback by continuing from a saved intermediate generation state.
Beihang University, Peking University · Sept 2026
- Big Techcs.CV
Foundation-model initialization improves feed-forward 4D hand-object reconstruction from video
4D-HOF uses conditional flow matching to refine foundation model estimates into accurate, stable 4D hand-object reconstructions.
NVIDIA, Texas A&M University · Oct 2026
- Industrycs.RO
Caching frozen model states makes forward-only robot adaptation much faster
VLA-ZO adapts vision-language-action models to distribution shifts 25x faster than baseline zeroth-order optimization.
UNIST · Oct 2026
- Top Universitycs.CV
Hierarchical implicit models improve compact brain MRI segmentation across sites
INR-based segmentation excels at low parameter budgets and limited augmentation, but U-Nets win with larger capacity.
University of Waterloo · Sept 2026
- Big Techcs.RO
Driving models can recognize hazards yet fail to brake
Driving VLAs can encode and describe hazards yet ignore them when planning; inference-time hazard guidance restores braking without retraining.
Politecnico di Milano, Stanford University · Oct 2026
- Big Techcs.CV
Correct image correspondences often in top-k but missed by top-1
Correct semantic correspondences are often in the top candidates but missed by greedy top-1 selection, which cycle-consistent re-ranking fixes.
Yale University, Amazon.com, Inc. · Nov 2026
- Big Techcs.AI
AI agents make corrections nearly twice as reliable by remembering mistakes
Self-Directed Repair makes agents learn from mistakes by storing lessons in lightweight LoRA adapters, reducing repeated errors.
Princeton University, Meta FAIR · Oct 2026
- Big Techcs.AI
Evolving environment beliefs helps LLM agents adapt without retraining
EvoBelief improves frozen LLM agents by validating, reusing, and revising falsifiable environment rules, including after hidden dynamics changes and during scientific discovery.
Tulane University, Amazon · Oct 2026
- Big Techcs.CV
Using a model's own past errors improves visual in-context learning
A model's own prediction errors can serve as counterfactual demonstrations to improve visual in-context learning.
NVIDIA · Oct 2026
- Academiccs.RO
Five human videos can adapt robot policies within five minutes
A 400M diffusion policy pretrained on 100 hours of egocentric human video learns new robot tasks from 5 demonstrations in under 5 minutes.
New York University · Oct 2026
- Big Techcs.LG
Language model agents evolve data pipelines to boost tabular foundation models
TabFM-Auto pairs a frozen tabular model with an LLM agent that iteratively refines the data pipeline, beating all competitors on 51 benchmarks.
Google Research, University of Southern California · Oct 2026
- Chinese Techcs.CV
Adaptive visual memory cuts multimodal reasoning costs without losing accuracy
Adaptive selection of fine-grained visual tokens during decoding improves multimodal reasoning accuracy and efficiency compared with one-shot compression.
City University of Hong Kong, Zhejiang University · Oct 2026
- Chinese Techcs.LG
Query-adaptive neuron steering improves frozen language models across benchmarks
Q-NeuS adaptively modulates neurons per query to boost LLM performance without retraining or task labels.
Northeastern University, Washington University in St. Louis · Oct 2026
- Top Universitycs.CV
Video models detect temporal order internally, then let the signal fade
Temporal information in VideoLLMs peaks in intermediate layers and fades before output, and reinjecting it improves temporal reasoning without training.
Korea Advanced Institute of Science and Technology (KAIST) · Oct 2026
- Big Techcs.CV
Training-free visual memory framework compresses video tokens while preserving event coherence
KeyRec uses a bounded visual memory with a recent cache and event bank to compress video for VLMs, using 10% of the token budget while improving accuracy.
University of Virginia, Google · Oct 2026
- Big Techcs.AI
FOCUS Compresses Agent Histories While Preserving Future Decisions
FOCUS cuts peak context by up to 48% and improves task success by up to 8.9 percentage points without training.
M365 Research, Microsoft · Sept 2026
- Top Universitycs.CV
Test-time method prevents drift in long video generation by keeping caching in-distribution
In-Distribution Forcing prevents drift in long video generation by aligning KV caching and conditioning with training configurations.
Seoul National University, Korea University · Oct 2026
- Chinese Techcs.CV
ReVision adapts frozen image generators to perception tasks with per-instance parameter updates
ReVision generates instance-specific low-rank updates from instructions to adapt frozen image generators for dense perception tasks.
Tencent Hy, Shanghai University of Finance and Economics · Oct 2026
- Top Universitycs.RO
Robot learns manipulation tasks from scratch in minutes with AI-generated 3D models
ZeroBot enables robot manipulation learning from scratch in minutes using image-to-3D generative models and parallel RL.
Imperial College London, Robotics and AI Institute · Oct 2026
- Academiccs.AI
Framework improves AI agents by patching common failure patterns
Vestrum turns execution trace failures into scoped harness changes that improve performance on held-out tasks without retraining the model.
University of Georgia · Oct 2026
- Big Techcs.AIcode
Adaptively choosing how to revise LLM skills outperforms fixed revision strategies
StraTune adaptively selects revision operators to improve self-evolving LLM skills, outperforming fixed operators across tasks.
The University of Texas at Austin, Amazon · Sept 2026
- Top Universitycs.AI
Bayesian reliability memory helps agents decide when to trust advice
BaRe-Mem uses Bayesian updates from past interactions to estimate advisor reliability, which then guides consultation decisions and response weighting.
DeCLaRe Lab, Nanyang Technological University, Peking University · Sept 2026
- Academiccs.OS
Prioritizing action-relevant context cuts LLM agent cache costs
ActKV retains 98.53% of FullKV accuracy using 25.98% peak KV cache memory, achieving 3.97x token and 3.58x task throughput.
University of Science and Technology of China, Suzhou Institute for Advanced Research, University of Science and Technology of China · Sept 2026
- Top Universitycs.RO
Robots can change task strategies when their controllers change
Robot policies can adapt to new controllers by observing how a few commanded actions actually move the robot, enabling behavioral transfer without explicit controller knowledge.
Stanford University · Oct 2026
- Top Universitycs.LG
Rewards Barely Drive Learning in Direct In-Context Reinforcement Learning
Direct in-context reinforcement learning is driven by trajectory imitation and surface form, not by reward semantics, and should be reframed as in-context learning.
Korea Advanced Institute of Science and Technology (KAIST) · Oct 2026
- Big Techcs.AI
Language models discover search algorithms that transfer across games
A multi-agent LLM system discovers modular game-playing search algorithms that outperform most MCTS baselines across more than 400 games and transfer to unseen domains.
Google DeepMind · Sept 2026
- Big Techcs.LG
Prompt-adaptive compression halves transformer KV-cache memory without dropping tokens
AttSVD compresses the KV cache by storing every token in a per-prompt low-rank subspace, matching dense performance using up to 50% memory.
Microsoft Applied Sciences Group (ASG) · Oct 2026
- Big Techcs.LG
Low-dimensional language-model steering supports more systematic black-box search
BOReFT learns a low-dimensional intervention manifold in a frozen LLM for Bayesian optimization of discrete black-box functions.
University of Massachusetts Amherst, New York University · Oct 2026
- Big Techcs.LG
Segment-level expert routing gives one model four inference budgets
Stepped MoE combines configurable active parameter budgets with segment-level expert routing, outperforming comparable dense models while retaining similar latency on memory-constrained devices.
Apple · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.