A Survey on Self-Improving Test-Time Intelligence: Feedback-Driven Adapting, Learning, and Scaling at Inference

Motivation

• Fixed train-then-infer models struggle with dynamic environments, distribution shifts, personalized demands, and varying task difficulty during deployment.Real-world inference therefore increasingly requires more than executing a frozen model once.Source: §1 · p. 3
• Deployment-time inference can exploit feedback through state updates, knowledge retrieval, tool interaction, adaptive computation, and multi-step reasoning or verification.These mechanisms motivate AI systems that improve during deployment rather than remaining static after training.Source: §1 · p. 4

Test-Time Learning and Adaptation

• Test-Time Learning (TTL), introduced by Sun et al., improves inference by updating model states with unlabeled test data, while Test-Time Adaptation (TTA) applies this idea through updates to layers, prompts, adapters, inputs, or parameters.TTA commonly targets distribution-shift robustness but can more broadly use test-time signals to improve prediction quality or task capability.Source: §1 · p. 4

Test-Time Scaling

• Test-Time Scaling (TTS) improves performance through extra inference-time computation, including repeated decoding [27], self-consistency, tool use [45] [97], search, and generative-model optimization [249].Planning and interaction can similarly improve embodied decisions without persistent parameter updates [181].Source: §1 · p. 4

Test-Time Intelligence

• TTI unifies learning and scaling as complementary routes to feedback-driven self-improvement, with learning emphasizing persistent or semi-persistent state change and scaling emphasizing additional computation without necessary state change.The routes can form hybrids through Scaling for Learning, which generates pseudo-labels, and Learning to Scale, which learns when to allocate more reasoning computation.Source: §1 · pp. 4, 6

Scope and Contributions

• The survey positions TTI as a broader framework connecting TTA, TTL, and TTS across machine learning, foundation models, generative models, embodied systems, and applications beyond language.It distinguishes adaptation as a subset of learning, scaling as related but not equivalent to learning, and hybrid systems as an important intersection while identifying transferable mechanisms and research gaps.Source: §1 · p. 6

2.1. Test-Time Intelligence

• TTI enables an AI system to improve during deployment by using test-time feedback, through test-time learning (TTL), test-time scaling (TTS), or both.The taxonomy includes model states such as norm layers, hidden states, prompts, adapters, and external memory.Source: §2.1 · p. 6, Figure 3 · p. 7
• TTL changes a persistent or semi-persistent writable model state through explicit learning, whereas TTS improves predictions by spending additional inference-time compute on search, verification, or tool use.Test-time adaptation is treated as a TTL special case focused primarily on robustness under distribution shift, while TTS typically leaves no lasting model change after the current inference.Source: §2.1 · p. 7
• TTL and TTS are interacting axes: learning can improve compute allocation and tool selection, while scaling can provide supervision for self-training or policy updates.Their interaction produces hybrid systems combining persistent self-improvement with dynamic inference-time computation.Source: §2.1 · p. 7
• TTI applies when deployment-time signals improve the active system immediately or cumulatively, whereas improvements deferred to fine-tuning, distillation, or training-data construction are training-time amortization.Ordinary transient states such as KV caches, sequential hidden states, and predefined diffusion latents are not TTL unless explicitly optimized or updated using feedback and reused by the deployed system.Source: §2.1 · pp. 7–8

2.2. A Unified View: Update and Compute

• The unified view represents TTI through Update and Compute: Update changes state using feedback, while Compute controls the amount and structure of inference-time computation.Update underlies TTL through parameter, prompt, normalization, latent, or memory changes; Compute underlies TTS through repeated decoding, extended reasoning, search, verification, or multi-agent interaction.Source: §2.2 · p. 8
• Classical TTL mainly occupies the state-update, lower-compute region, TTS the higher-compute, nonpersistent-update region, and hybrids the intersection by coordinating both mechanisms.Hybrid methods may use search or self-consistency to improve learning signals, or learn when and how to allocate additional inference-time resources.Source: §2.2 · p. 8

• TTL updates a deployed model’s writable state from test inputs and, when available, external feedback, using the updated state during the same deployment process.Source: §3 · p. 8
• Because deployment data can differ from training data through sensor, environmental, domain, or temporal changes, frozen models inevitably degrade, motivating TTL beyond adaptation toward memorization, personalization, and self-improvement.Source: §3 · p. 8

Problem Setting

• Test-time learning updates writable state using deployment data and optional external feedback, with the state encompassing inputs or representations, parameters or statistics, auxiliary modules, and external memory.The writable state changes during the same deployment process and is then used for inference.Source: §3.1 · p. 9
• TTL covers both distribution-shift settings where pS ≠ pT and same-distribution settings where it enables adaptive memorization and self-improvement.Updates may optimize a test-time objective, adjust running statistics, modify memory, or refine inputs or latent representations.Source: §3.1 · p. 9

Design Dimensions

• The design space is organized by feedback signal, update target, temporal horizon, and objective scope.These dimensions respectively describe what drives updates, which writable state changes, how information accumulates from individual samples to continual deployment, and whether the goal is alignment, memory, personalization, or self-improvement.Source: §3.1 · p. 9
• Methods are described across these axes, allowing compound methods to appear under multiple categories while locating each mechanism where it is most informative and cross-referencing its other roles.Sections 3.2–3.4 emphasize feedback, update targets, and temporal attributes, whereas Sections 3.5–3.6 classify methods by primary objectives.Source: §3.1 · p. 9

3.2. Feedback Signals for Test-Time Learning

• Feedback signals drive writable-state updates; without labels, test-time learning must construct surrogate objectives correlated with downstream performance from available deployment information.Signal choice is consequential because each family makes different assumptions about reliable supervision.Source: §3.2 · p. 9

3.2.1. Entropy and Confidence-Based Signals

• Entropy and confidence signals offer inexpensive adaptation or update gating, but can reinforce confident mistakes, discard useful hard samples, or collapse under calibration and distribution shifts.Tent minimizes prediction entropy, while EATA, SAR, DeYO, DELTA, SHOT, and related methods improve filtering, calibration, or collapse resistance.Source: §3.2 · p. 9, §3.2.1 · pp. 10–11

3.2.2. Consistency and Augmentation-Based Signals

• Consistency signals improve robustness by enforcing agreement across augmented views, model branches, partial inputs, or retained feature statistics, but require valid transformations, stable targets, or representative references.Ada-ReAlign adapts evolving references for non-stationary streams, though online-reference errors can accumulate.Source: §3.2.2 · pp. 11–12

3.2.3. Reconstruction and Self-Supervision Signals

• Reconstruction and other self-supervised signals provide dense input- or feature-grounded feedback, yet depend on training-prepared auxiliary components and correlation between pretext objectives and downstream tasks.Methods include TTA-AE, TTT-MAE [88], TTT, TLM [128], TTT++ [240], and auxiliary–main task alignment approaches such as SR-TTT.Source: §3.2.3 · p. 12

3.2.4. Pseudo-Label and Self-Training Signals

• Pseudo-label and self-training methods can derive supervision from adapted-model predictions or external teachers, but self-generated targets may reinforce incorrect predictions, while teachers may transfer bias or domain mismatch.Source: §3.2.4 · p. 13

3.2.5. External Feedback from Tools, Environments, or Users

• External feedback from environments, tools, and users is more task-aligned but introduces interfaces, latency, permissions, human effort, evaluator bias, and sparse or delayed supervision.Environment outcomes, executable tools, model evaluators, binary judgments, queried labels, and preferences provide progressively different task-alignment and feedback-bandwidth tradeoffs.Source: §3.2 · p. 10, §3.2.5 · p. 14

3.3. What Is Updated at Test Time?

• The update target is the central design choice in TTL, yielding five categories whose scope balances adaptation capacity, stability, and efficiency rather than forming a strict strength hierarchy.The categories are input adaptation, partial-parameter updates, auxiliary parameter updates, backbone or full-parameter updates, and external memory or cache updates.Source: §3.3 · pp. 14–15, Table 2 · p. 15

3.3.1. Input Adaptation

• Input adaptation modifies test inputs instead of model parameters, using camera, diffusion, or frequency-domain interventions to reduce domain mismatch while avoiding parameter access.These methods can operate independently per sample, but iterative restoration adds latency and may remove task-relevant content; frequency calibration avoids generative restoration but depends on suitable priors.Source: §3.3.1 · pp. 15–16

3.3.2. Partial Parameter Updates

• Partial parameter updates localize changes to normalization statistics, selected channels or layers, or adaptation-critical parameters, preserving other weights while keeping test-time cost tractable.Normalization methods are inexpensive but depend on representative target statistics and stable streams, while selective updates require reliably identifying which parameters should adapt.Source: §3.3.2 · p. 16

3.3.3. Auxiliary Parameter Updates

• Auxiliary parameter updates freeze the pretrained backbone and add prompts, adapters, or low-rank modules, trading reversibility and storage efficiency against restricted capacity and architecture-dependent placement.Prompt methods strengthen evidence, reuse history, or constrain optimization, while low-rank parameterization, task-structured modules, and routed compositions determine how adaptation capacity is distributed.Source: §3.3.3 · pp. 16–17

3.3.4. Backbone and Full-Parameter Updates

• Backbone-only and full-model updates provide broader correction capacity, but require reliable feedback because representation drift, decision-boundary degradation, forgetting, and collapse can affect subsequent predictions.Rollback and stabilization mechanisms include stochastic restoration in CoTTA, adaptive re-initialization in ABR, consistency-based backtracking in PLATO-TTA, and parameter-shared models in AWMC.Source: §3.3.4 · pp. 17–18

3.3.5. External Memory and Cache Updates

• External memory and cache updates retain deployment-time knowledge outside the model through class prototypes, instance retrieval, or structured experience, enabling gradient-free reuse but introducing storage, retrieval, contamination, staleness, and privacy risks.Prototype memories are compact but can be distorted by imbalance or incorrect assignments; instance memories preserve richer structure but require more computation and remain sensitive to noisy neighbors, while semantic memories support later reasoning or action.Source: §3.3.5 · p. 18

3.4. Temporal Horizons of Test-Time Learning

• Temporal design separates the supervision horizon—from a single input, token, or interaction to a batch, history, or trajectory—from the persistence horizon governing state lifetime.A method may update from one observation and retain the result across a stream, or aggregate a trajectory within an episode and then reset.Source: §3.4 · p. 19

3.4.1. Supervision Horizon: Single- vs. Multi-Sample Learning

• Single-sample supervision enables immediate, low-buffer updates through prediction- or structure-derived signals, but signals may be noisy and repeated optimization costly.MEMO and TPT use one input, while TTT, MAE-TTT [88], and TTT-Layer use pretext, reconstruction, or token-level self-supervision.Source: §3.4.1 · p. 19
• Multi-sample supervision aggregates related observations across batches, histories, or interaction sequences, improving evidence reliability while introducing coherence, buffering, staleness, or ordering requirements.Tent, TTT++ [240], and ActMAD aggregate batches; RoTTA, SAR2, ViTTA, BayesTTA [66], TT-VLA, and ReflectivePlanning [124] use histories, temporal evolution, or feedback sequences.Source: §3.4.1 · p. 19

3.4.2. Persistence Horizon: Episodic and Continual Learning

• Episodic persistence resets writable state after bounded deployment units, limiting cross-unit error propagation while permitting within-episode context reuse.TTT, MEMO, TPT, and MAE-TTT [88] commonly restore source state per input, whereas TTT-Layer retains fast-weight memory within a sequence and Tent may reset or retain state independently of batch updates.Source: §3.4.2 · p. 20
• Continual persistence reuses experience across deployment streams for adaptation, personalization, and planning, but requires controls against drift, contamination, staleness, excessive growth, and cross-context leakage.Anchoring, reliability-based admission, modular state, and external memory regulate retention in CoTTA, EATA, SAR, RoTTA, AETTA, CoLA, BECoTTA, Cheatsheet, PhysMem, and ReflectivePlanning [124].Source: §3.4.2 · p. 20

3.5.1. Domain Shift and Distribution Alignment

• Domain shift changes p(x), p(y), or p(x | y); TTL minimizes source–target representation divergence, with covariate and label shifts often occurring together.Covariate shift preserves p(y | x), whereas label shift preserves p(x | y); real deployments may combine both.Source: §3.5.1 · pp. 20–21

3.5.2. Test-Time Training and Fully Test-Time Adaptation

• TTT prepares an auxiliary objective during source training, whereas FTTA adapts off-the-shelf models from test-time signals, trading preparation cost against exposure to unreliable feedback.TTT benefits from task-aligned auxiliary objectives; FTTA broadens usable signals through entropy, consistency, and pseudo-label objectives but risks confirmation bias, drift, and collapse.Source: §3.5.2 · pp. 21–22

3.5.3. Source-Free, Online, and Continual Adaptation

• Source-free, online, and continual protocols improve deployment practicality but remove anchors, restrict adaptation evidence, and increase risks of drift, collapse, catastrophic forgetting, and error accumulation.Offline adaptation enables dataset-level statistics and longer optimization, while continual methods use sample selection, calibrated and anchored updates, and multi-modeling to improve stability.Source: §3.5.3 · pp. 22–23

3.5.4. Forward-only and Gradient-Free Adaptation

• Forward-only TTL replaces backpropagation with derivative-free optimization or direct state calibration, addressing inaccessible models and memory overhead while limiting update flexibility.For example, ZOA achieves higher learning efficiency than CMA-ES with two forward passes per sample, and EVA-0 stabilizes this setting using scale-invariant, anchor-guided, symmetric perturbations.Source: §3.5.4 · pp. 23–24

3.5.5. Theoretical Understanding of Test-Time Adaptation

• Test-time adaptation theory provides no universal guarantee: improvements require proxy–task alignment and mechanism-specific assumptions, while unreliable feedback or uncontrolled updates can cause degradation and collapse.Bounds cover settings such as aligned auxiliary gradients, memory or labeled feedback, gradual shifts, and dynamic-regret streams; protection requires informative signals, controlled updates, and harmful-shift detection.Source: §3.5.5 · p. 25

3.6. Beyond Adaptation: Memory, Personalization, and Self-Improvement

• Together, these beyond-adaptation families emphasize whether deployment information persists as memory, modifies the agent or improvement process, specializes a user or context, or supplies corrective feedback.The framework distinguishes evolving memory, personalization, environment- or task-specific learning, interactive improvement, and recursive self-improvement by persistence, specialization target, and external feedback source.Source: §3.6 · pp. 25–26

3.6.1. Test-Time Learning with Evolving Memory

• Evolving memory stores, organizes, and retrieves deployment observations or feedback so later predictions and actions can reuse context and accumulated experience.TTT-Layer changes quadratic long-context processing O(n^2) into recurrent linear sequence complexity O(n), while later systems manage retention, forgetting, and reusable strategies across tasks and sessions.Source: §3.6.1 · p. 26

3.6.2. Recursive Self-Improvement

• Recursive self-improvement closes a loop in which systems propose, evaluate, retain, and reuse modifications to solutions, agents, or the process generating future improvements.FunSearch, AlphaEvolve [273], ADAS [130], and self-modifying agents progressively enlarge the writable target, while Hyperagents [476] and MetaSkill-Evolve [409] modify the improvement process itself; verification requires held-out evaluation, regression testing, version isolation, and rollback.Source: §3.6.2 · p. 27

3.6.3. Personalization and User-Specific Learning

• Personalization specializes pretrained models to individual users, patients, or instances whose characteristics are not adequately modeled by population-level training.Applications include user-preference language generation, speaker and acoustic personalization, human-centered calibration, patient-specific medical modeling, personalized brain-computer interfaces, protein adaptation, and patient-aware clinical time series.Source: §3.6.3 · pp. 27–28

3.6.4. Environment-Specific and Task-Specific Specialization

• Environment- and task-specific specialization tailors general-purpose models to recurring deployment conditions, often using limited context- or task-specific data without task-specific fine-tuning data.Examples span robotic geometry, dynamics, and task feedback; industrial and scientific operating conditions; and foundation-model prompt, activation, retrieval, or input-learning specialization for downstream domains.Source: §3.6.4 · p. 28

3.6.5. Interactive Improvement from User Feedback

• Interactive improvement uses direct human feedback to correct active systems or align outputs with deployment-specific intent, ranging from inexpensive evaluations to structured corrections.Binary judgments reduce annotation effort but weakly specify revisions, whereas clinician corrections, scribbles, clicks, and 2D edits provide targeted supervision at greater interface and user-effort cost; verifier-only selection belongs to TTS, while rollout-derived updates belong to test-time reinforcement learning.Source: §3.6.5 · p. 29

4.1. Problem Formulation

• Test-time scaling keeps pretrained parameters θ fixed and varies an inference-time budget c, including samples, reasoning steps, search depth, verification rounds, or tool calls.The baseline procedure is recovered at c = c0, while scaling controls additional computation within the active inference process.Source: §4.1 · p. 29

4.2. Scaling Compute at Inference Time

• Inference compute can extend one trajectory, generate and compare complete candidates, explore partial-state search spaces, or adapt budgets to input difficulty and reasoning stage.Longer reasoning suits incremental improvement, repeated sampling suits diverse complete answers, and search suits evaluable partial states, lookahead, or backtracking.Source: §4.2 · p. 30

4.2.1. Longer Reasoning and Deeper Inference

• Longer or deeper inference can improve reasoning through intermediate steps, revision, latent iterations, structured decomposition, or algorithmic search, but excessive length can propagate errors and cause overthinking.Wu et al. report an inverted U-shaped relation between reasoning length and acc.Source: §4.2.1 · pp. 30–31

4.2.2. Best-of-N Sampling and Repeated Decoding

• Best-of-N sampling broadens candidate coverage, but improves final accuracy only when candidates are sufficiently diverse and the selector or verifier is reliable.Methods use learned verification, output-distribution confidence, pairwise tournaments, reward regularization, or early rejection to select among candidates.Source: §4.2.2 · p. 31

4.2.3. Self-Consistency and Consensus Mechanisms

• Consensus methods select frequent or weighted answers efficiently, yet correlated samples, ambiguous equivalence, or excessive calls can turn additional votes into confident shared errors.Self-consistency, CISC, Soft-SC, and DSC respectively vary aggregation or adapt sampling to difficulty.Source: §4.2.3 · p. 32

4.2.4. Search-Based Inference and Planning

• Search-based inference preserves relations among partial states, enabling lookahead, revision, backtracking, and evaluator-guided expansion for planning or combinatorial problems.Its benefits can disappear when noisy evaluators discard promising branches or branching, frontier memory, and sequential control exhaust the budget.Source: §4.2.4 · pp. 32–33

4.2.5. Adaptive Computation and Dynamic Budgeting

• Adaptive computation allocates reasoning depth, sampling, or search according to estimated difficulty, potentially reducing average cost or allowing smaller models to outperform larger ones, but poor estimates waste computation or stop early.Compute-optimal scaling, DSC, Meta-Reasoner, s1, and LCPO [3] illustrate adaptive allocation and length control.Source: §4.2.5 · p. 33

4.2.6. Theoretical Understanding of Test-Time Scaling

• Additional computation yields conditional gains: sampling addresses estimation error, while reliable verification and allocation are needed because gains can saturate or worsen with poor coverage, reward hacking, or weak evaluators.Theoretical and empirical results show that more computation is not inherently beneficial and that comparable guarantees remain limited beyond language reasoning.Source: §4.2.6 · pp. 33–34

4.3. Scaling with External Resources

• External resources broaden inference beyond internal computation by providing unavailable evidence, exact execution, independent assessment, or additional reasoning diversity.Table 4 compares retrieval, tools, verifiers, and multi-agent inference according to the bottlenecks they address and their deployment tradeoffs.Source: §4.3 · p. 34

4.3.1. Retrieval and External Knowledge

• Retrieval methods adapt evidence acquisition through reflection, uncertainty triggers, iterative reasoning loops, complexity-based depth control, structured search, and cross-path verification, but face evidence quality and latency limits.Self-RAG [7], FLARE, IRCoT, Adaptive-RAG, Search-o1, Search-R1 [158], and MIRAGE exemplify these mechanisms.Source: §4.3.1 · p. 35

4.3.2. Tool Use and Program Execution

• External tools delegate unreliable operations to executors and use observations or execution feedback to answer, validate, revise, switch strategies, or synthesize executable tools.ReAct, Program of Thoughts (PoT) [45], START, CRITIC [97], XoT, and Test-Time Tool Evolution (TTE) [243] illustrate these patterns; benefits depend on reliable interfaces and correct invocation.Source: §4.3.2 · p. 35

4.3.3. Verifiers, Critics, and Re-Rankers

• Verifiers and critics improve candidate selection, refinement, and search guidance through outcome scores, process feedback, verifier reasoning, textual critiques, or aggregated judgments, but cannot recover absent solutions and may mis-rank errors.Verifier-guided BoN [60], process reward models, AlphaMath [35], GenRM, Critique-out-Loud (CLoud) [5], and Multi-Agent Verification (MAV) represent these mechanisms.Source: §4.3.3 · pp. 35–36

4.3.4. Multi-Agent Inference and Collaborative Reasoning

• Multi-agent inference scales proposals, expertise, specialization, tools, interaction, and revision through voting, layered aggregation, consensus, debate, structured networks, and iterative sharing, but agent count alone does not ensure improvement.Agent Forest, Mixture-of-Agents (MoA), ReConcile [41], Multi-Agent Debate (MAD), MacNet, METAL, and TUMIX [49] show gains from complementary information or error patterns, while correlated errors, premature convergence, and communication costs can degrade performance.Source: §4.3.4 · p. 36

4.4.1. Reasoning Quality

• Scaling improves reasoning by exposing deeper and broader trajectories, then structuring, evaluating, and adaptively allocating computation across alternatives.Chain-of-Thought increases depth, while self-consistency and repeated sampling increase breadth; tree or graph search, supervision, verifiers, and adaptive allocation organize and filter trajectories.Source: §4.4.1 · p. 36
• Longer reasoning alone is insufficient: scaling helps when extra computation yields diverse, task-relevant trajectories and reliable evaluation, but can amplify redundant or incorrect reasoning otherwise.Source: §4.4.1 · p. 36

4.4.2. Reliability and Calibration

• Additional computation supports verification, candidate selection, and calibration through tool-based checking, retrieval and reflection, confidence-aware aggregation, and agreement-based uncertainty signals.CRITIC examines and refines outputs, Self-RAG decides when evidence or revision is needed [7] [97], and confidence-aware aggregation can weight candidates, stop early, or allocate samples.Source: §4.4.2 · p. 37
• Selection and calibration are distinct: consensus may improve candidate choice without revealing shared confident errors, so reliable scaling requires both effective aggregation and informative uncertainty.Source: §4.4.2 · p. 37

4.4.3. Decision Quality in Control and Planning

• In control and planning, scaling organizes, compares, and updates candidate decisions through decomposition, hierarchical execution, trajectory evaluation, environment interaction, and world models.ADaPT decomposes failed tasks, Plan-and-Act separates planning from execution [82], and ARMAP evaluates longer-term trajectory value [54].Source: §4.4.3 · p. 37
• These planning benefits require meaningful decomposition, trajectory-level feedback, or faithful environment models; inaccurate signals can make additional search optimize the wrong action path.Source: §4.4.3 · p. 37

4.5.1. High Computational Cost

• Repeated sampling, search, and verification increase model calls and serving overhead, while compute-optimal allocation reduces waste but cannot eliminate generation and evaluation costs [27].Tree search can be substantially slower than simpler baselines while yielding limited incremental gains in some settings [53].Source: §4.5.1 · p. 38

4.5.2. No Persistent Improvement by Default

• Pure scaling usually improves only the current inference episode, so later queries require repeated search unless results are retained in parameters, memory, or offline learning [134] [227].Persistence requires an additional retention mechanism; pure scaling alone does not provide it.Source: §4.5.2 · p. 38

4.5.3. Diminishing Returns and Instability

• Additional computation can saturate through redundant candidates and counterproductive reasoning, while unreliable self-critique and poorly coordinated agents can amplify errors and variance [203].Scaling remains useful only while generation is informative, evaluation distinguishes useful trajectories, and allocation stops before marginal gains are exhausted.Source: §4.5.3 · p. 38

4.5. Limitations of Pure Scaling

• Scaling is worthwhile only when reliable marginal benefits exceed latency and resource costs and its episode-local or persistent effects fit the intended use.These limitations connect computational cost, persistence, and diminishing returns to deployment decisions.Source: §4.5.3 · p. 38, §4.5 · p. 38

5. The Intersection of Learning and Scaling

• Across these relationships, additional computation can either become persistent learning, be learnedly allocated or routed, or close a deployment-time feedback loop that changes subsequent inference.Figure 8 summarizes the taxonomy: scaling for learning produces reusable signals, learning to scale controls inference computation, and joint systems update the active system during deployment.Source: §5 · pp. 38–39, Figure 8 · p. 39

5.1. Scaling for Learning

• Scaling for learning converts generation, verification, or search into reusable supervision or parametric capability, usually through offline training rather than TTI.Consensus and self-generated targets offer cheap but coarse signals, verifier-derived feedback improves credit assignment at higher evaluation cost, and search-derived trajectories provide richer supervision but can inherit evaluation and search biases.Source: §5.1 · p. 39, §5.1.1 · p. 39, §5.1.2 · p. 40

5.2. Learning to Scale

• Learning to scale trains controllers or inference policies to allocate computation, route among mechanisms, or exploit additional reasoning, retrieval, tools, and search effectively.Adaptive allocation trades control overhead against calibration and monitoring costs; routing selects internal reasoning, external mechanisms, or models; learned scaling behaviors prepare self-correction or search policies for deployment.Source: §5.2 · pp. 40–41, §5.2.1 · p. 41, §5.2.2 · p. 41

5.3. Joint Learning-and-Scaling Systems

• Joint learning-and-scaling systems write inference-derived feedback into the active model’s parameters, modules, or state and reuse the update within the same deployment process.This distinguishes them from pure scaling, which leaves no learned state, and from later offline training; the section organizes them into online reinforcement learning and self-generated adaptation data.Source: §5 · p. 39, §5.3 · p. 42

5.3.1. Test-Time Reinforcement Learning

• Test-time reinforcement learning explores candidate solutions, constructs online rewards, and updates the active model, with consensus, verified, stabilized, or task-grounded feedback providing different reliability and access-cost tradeoffs.TTRL, TTRV [333], T3RL [217], DARE [77], SCOPE, DDRL, TTT-Discover [461], and Alpha-RTL [500] exemplify these reward constructions; reliability is greatest when rewards remain correlated with correctness.Source: §5.3.1 · p. 42

5.3.2. Self-Synthesized Supervision

• Self-synthesized supervision generates adaptation data during deployment and updates the active model using direct self-editing, instance-conditioned examples, or reflection-guided curricula.SEAL, TT-SI, QueST [339], MASS [170], TTSR [117], and TTCS [431] reduce dependence on external labels, but require reliable filtering because synthetic errors can be reinforced.Source: §5.3.2 · p. 42

6.1. Vision Perception

• Vision applications exploit shifts in observations and ordered multimodal evidence for adaptation, while video methods must prevent errors from accumulating across streams.Conventional test-time learning addresses classification, segmentation, and detection; video methods use temporal statistics, self-supervision, prompt tuning, cached evidence, and modality collaboration.Source: §6.1 · p. 43

6.2. Generative Models

• Generative-model applications refine restoration, synthesis, video, and 3D outputs through model-, input-, observation-, or trajectory-level test-time intervention without retraining the backbone.Video methods target motion, appearance, temporal coherence, and compositional alignment; 3D methods use prior-guided generation or observation-guided reconstruction, whose benefits depend respectively on 2D-prior alignment or measurement completeness.Source: §6.2 · p. 44, §6.2.1 · p. 44, §6.2.2 · p. 44

6.2.4. Inference-Time Scaling

• Inference-time scaling makes diffusion denoising trajectories searchable through noise optimization, branching, selection, steering, and reflection, but requires reliable evaluators and incurs additional sampling cost.Scaling also extends to autoregressive image generation, where token-entropy confidence prunes trajectories and schedules guidance.Source: §6.2.4 · p. 45

6.3. Language and Multimodal Models

• Language and multimodal applications emphasize specialized spatial and speech settings, using verification, imagined viewpoints, adaptive evidence generation, and audio-specific adaptation to address weak supervision and sequential dynamics.Speech methods progress from entropy minimization and single-utterance adaptation to sequence-level, continual, dynamic, and backpropagation-free prompt adaptation; E-BATS reduces memory while maintaining comparable efficacy.Source: §6.3.1 · pp. 45–46, §6.3.2 · p. 46

6.4. Embodied AI and Robotics

• Embodied TTI intervenes in the perception–planning–action loop through policy adaptation, reusable skill improvement, and planning-time search, verification, and selection over grounded behavior.Policy updates may target perception, planners, controllers, or the action policy; skill methods expand or refine reusable capabilities, whereas planning-time scaling improves current decisions without persistent agent updates.Source: §6.4 · p. 46, §6.4.1 · p. 47, §6.4.3 · pp. 47–48

6.5. Agentic AI

• Agentic systems use environment observations, state transitions, collaboration experience, runtime monitoring, and self-assessed uncertainty to drive test-time adaptation and self-improvement.GTTA [33], MAS-on-the-Fly [232], and TT-SI illustrate environment-specific alignment, multi-agent intervention, and targeted updates for difficult cases.Source: §6.5 · p. 48

6.6. Healthcare and Personalized AI

• Healthcare applications distinguish environment-level adaptation from patient-specific personalization, requiring uncertainty controls, repeated patient observations, reliability safeguards, and privacy-aware deployment.Federated test-time personalization reduces raw-data centralization but does not formally guarantee privacy for information encoded in updated model states.Source: §6.6 · pp. 48–49

7.1. Open Challenges

• TTI needs reliable, broadly applicable objectives and stronger long-horizon stability because noisy unsupervised signals can cause model collapse in open-world deployment.Promising mitigations include unreliable-sample filtering [133], sharpness-aware optimization, feature-diversity regularization, and online risk monitoring.Source: §7.1 · p. 49
• TTI remains costly and narrowly evaluated, motivating near-single-forward-pass efficiency and realistic benchmarks that jointly measure accuracy, latency, memory, compute, calibration, and stability.Current gaps include mixed, evolving, multimodal, personalized, temporal, and interactive shifts, plus missing matched test-time budgets and comparable feedback access.Source: §7.1 · pp. 49–50
• Deployment-ready TTI also requires self-calibrating hyperparameters and a unified theory explaining when feedback, writable state, and extra computation improve performance rather than accumulate errors.Theory should cover proxy-task alignment, update stability, sampling and verification, budget allocation, state-versus-computation choices, persistence of gains, regret, and risk.Source: §7.1 · p. 50

7.2. Future Directions

• Extending test-time learnable hidden states into scalable long-term memory architectures is a promising direction for handling increasingly important long contexts.Follow-up work improves objectives, memory capacity and reliability, transferability, and video extension [86] [211] [231] [359] [361] [408], while MGTTA [71] memorizes historical gradients for faster adaptation.Source: §7.2 · pp. 50–51
• TTI needs systematic adversarial evaluation and defenses because attacks can poison adaptive state or manipulate candidate diversity and evaluators, causing persistent degradation.Protocols should specify attacker access, budget, and timing, while defenses should detect unreliable feedback, constrain harmful state changes, and enable recovery.Source: §7.2 · p. 51
• Black-box adaptation of closed APIs remains underexplored, with BETA offering an early online unsupervised step but limited to classification-based vision APIs.Many prior approaches are supervised offline or gray-box because they require prompt tuning or source-domain statistics [75].Source: §7.2 · pp. 51–52
• Future TTI systems can improve diffusion generation through feedback on intermediate states and adaptive computation allocation, while human-interactive TTI can use sparse corrections, preferences, or instructions for personalized and controllable improvement.Key requirements are robust feedback, focused exploration or refinement, efficient incorporation of sparse human signals, low interaction cost, and stable improvement.Source: §7.2 · p. 52
• TTI could support scientific discovery and deep research by combining inference-time adaptation with tools, simulators, retrieval, optimization, human expertise, and full research workflows [56] [461] [208].These systems could search for exceptional solutions and advance models beyond current SOTA, but should assist researchers rather than generate papers for direct submission.Source: §7.2 · p. 52

• TTI unifies test-time learning, test-time adaptation, and test-time scaling through state updates and inference-time compute.The survey reviews learning-based adaptation, inference-time scaling, and their intersection as parts of this framework.Source: §8 · p. 53
• The survey covers TTI applications across vision, language, multimodal learning, generative models, robotics, and healthcare.Source: §8 · p. 53
• Key challenges include reliable test-time objectives, long-horizon stability, efficiency, realistic evaluation, theoretical foundations, and safety risks.The survey characterizes TTI as an emerging area despite rapid progress.Source: §8 · p. 53
• The survey offers a foundation and roadmap for developing adaptive, scalable, and self-improving AI systems at test time.Source: §8 · p. 53