PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Shuaicheng Niu, Guohao Chen, Yaofo Chen, Zhiquan Wen, Jinwu Hu, Zeshuai Deng, Deyu Chen, Shuhai Zhang, Renjie Chen, Zihao Lian, Shoukai Xu, Gang Dai, Yunbei Zhang, Wei Luo, Yifan Zhang, Mingkui Tan, Cheng Deng

arXiv:2609.01679v1cs.LG

TL;DR

AI systems increasingly need to improve during deployment because fixed train-then-infer models face dynamic and personalized environments. This survey introduces feedback-driven Test-Time Intelligence as a unified framework for state updates and additional inference-time computation, connecting test-time adaptation, learning, and scaling across applications. It concludes that TTI is an emerging area with unresolved challenges in reliability, stability, efficiency, and evaluation.

  • Problem

    Fixed train-then-infer models are increasingly mismatched with dynamic, distribution-shifting, and personalized deployment settings.

  • Method

    The survey organizes test-time learning, adaptation, and scaling around test-time feedback, state updates, and additional inference-time computation.

  • Results

    The survey provides a unified perspective connecting TTA, TTL, and TTS across methods and applications in vision, language, multimodal learning, generative models, robotics, and healthcare.

  • Takeaways & Limitations

    TTI supplies a common language for distinguishing state updates, added computation, and hybrid systems while revealing transferable mechanisms and research gaps.

  • Takeaways & Limitations

    Pure test-time scaling adds latency and resource cost, usually provides episode-local gains, and may eventually become redundant or unstable.

Abstract

from arXiv · show

The ability of AI systems to improve their behavior during deployment is becoming increasingly important. As inference moves beyond the static execution of a fixed trained model, a growing body of work studies how models can refine their behavior on the fly by exploiting test-time information and additional computation. These developments have largely evolved along two directions: methods that modify the model's state using test-time signals, and methods that improve predictions through extra inference-time resources such as more sampling and tool use. However, these directions are often studied in separate communities with different terminology, making their connections harder to see. In this survey, we present feedback-driven Test-Time Intelligence (TTI) as a unified perspective for understanding such deployment-time improvement. We use this view to relate test-time adaptation, test-time learning, and test-time scaling, highlighting both their distinctions and their growing overlap in hybrid systems. This unified framework helps connect previously fragmented ideas and provides a clearer conceptual foundation for studying inference-time self-improvement. We review major methodological paradigms, representative applications, and open challenges across vision, language, multimodal learning, generative models, robotics, and healthcare. Our goal is to provide a coherent foundation and research roadmap for the study of self-improving AI systems at test time.

1. Introduction

The survey frames deployment-time improvement as feedback-driven Test-Time Intelligence, unifying state updates and additional inference-time computation across previously separate research communities. It relates test-time learning, adaptation, and scaling while reviewing their applications and hybrid interactions.

  • Motivation: Dynamic deployment environments challenge the traditional assumption that a frozen model can reliably handle every test-time scenario.Models increasingly face distribution shifts and personalized settings after deployment.
  • Motivation: Inference-time self-improvement can use feedback to update internal state, retrieve knowledge, interact with tools, allocate computation, or refine outputs through reasoning and verification.These mechanisms move inference beyond static execution of a fixed trained model.
  • Conceptual framing: Test-Time Learning improves behavior by updating model states, whereas Test-Time Scaling improves performance through additional inference-time computation without necessarily changing states.Examples include adaptation of normalization layers, prompts, adapters, inputs, or parameters, alongside repeated decoding, self-consistency, tools, and search.
  • Conceptual framing: Feedback-driven Test-Time Intelligence defines deployment-time improvement through test-time feedback for state updates or additional computation.The framework treats learning and scaling as complementary routes toward intelligent inference.
  • Scope and contributions: The survey unifies TTA, TTL, and TTS, clarifies their relationships, and reviews paradigms, applications, and open challenges across multiple AI domains.It positions adaptation as a subset of learning and scaling as overlapping with, but not equivalent to, learning.
  • Scope and contributions: The unified perspective addresses fragmented terminology by connecting adaptation, inference-time scaling, embodied search and planning, and generative-model optimization.Its update–compute view distinguishes improvement from state updates, additional computation, or both.

2. Conceptual Foundations

Test-Time Intelligence (TTI) frames deployment-time self-improvement as feedback-driven inference through two complementary mechanisms: updating writable state and allocating additional computation. Test-time learning and scaling can also interact, producing hybrid systems that combine persistent changes with dynamic inference resources.

  • TTI defines self-improvement during deployment as using test-time feedback to adapt, refine, or enhance system behavior.
  • Test-time learning improves predictions through persistent or semi-persistent writable-state changes, whereas test-time scaling improves predictions through additional inference-time computation.
  • The TTL–TTS boundary is not strict: learning can guide when or how much to scale, while scaling can provide supervision for learning.
  • Ordinary transient inference states such as KV caches, sequential hidden states, and diffusion latents are excluded from TTL unless test-time feedback explicitly updates reusable writable state.
  • The unified framework organizes methods around update, which changes state using feedback, and compute, which determines the amount and structure of inference-time computation.
  • Classical learning methods primarily occupy the state-update region, scaling methods the extra-computation region, and hybrid methods combine both mechanisms.

3. Test-Time Learning

Test-time learning updates a deployed model’s writable state using test inputs and, when available, external feedback, then reuses that updated state during the same deployment process. Its scope includes distribution-shift adaptation as well as memorization, personalization, and self-improvement.

  • TTL updates writable state from test inputs and optional external feedback, using the updated state for inference within the same deployment process.

3.1. Problem Settings and Design Dimensions

The survey formalizes TTL as a deployment-time state transition and organizes its design space by feedback, update target, temporal horizon, and objective scope. These dimensions cover both distribution-shift settings and adaptive behavior when source and target distributions match.

  • TTL starts from a trained model with fixed parameters, then updates writable deployment state using test inputs and optional external feedback.
  • The update operator may optimize a test-time objective, apply running statistics or closed-form updates, modify memory, or refine inputs and latent representations.
  • Feedback signal describes the information driving updates, including entropy, consistency, pseudo-labels, and human, tool, or environmental feedback.
  • Update target specifies the writable component, ranging from inputs and normalization statistics to auxiliary modules, full parameters, and external memory.
  • Temporal horizon ranges from single-sample or batch adaptation to episodic or continual learning across a deployment lifetime.
  • Objective scope distinguishes distributional alignment, memory augmentation, personalization, and iterative self-improvement.

3.2. Feedback Signals for Test-Time Learning

Test-time learning derives surrogate supervision from internal signals, self-supervision, pseudo-labels, consistency, or external feedback because deployment often lacks ground-truth labels. Each signal family trades off cost, alignment, reliability, and assumptions about calibration, transformations, auxiliary objectives, or initial predictions.

  • When ground-truth labels are unavailable, TTL constructs surrogate objectives from test-time information that should correlate with downstream task performance.
  • Entropy and confidence signals are inexpensive with small parameter subsets but require calibration and can reinforce confident mistakes or collapse under severe shifts.
  • Consistency methods use extra forward passes and assume selected transformations preserve validity, reducing dependence on any single prediction.
  • Reconstruction and other pretext objectives provide dense input-grounded feedback but may optimize structure weakly related to the downstream task.
  • External tool, environment, or user feedback can be more task-aligned, while requiring interfaces, latency budgets, permissions, or human effort.
  • Pseudo-label and self-training methods derive supervision from adapted-model predictions, prototypes or neighbors, prediction consensus, or external models, but can suffer confirmation bias.

3.3. What Is Updated at Test Time?

Test-time learning organizes methods by which component or deployment state is updated, balancing adaptation capacity, stability, and computational burden. The taxonomy spans input changes, parameter updates, auxiliary modules, full-model updates, and external memory.

  • Test-time update targets comprise input adaptation, partial parameter updates, auxiliary parameters, backbone or full-parameter updates, and external memory or cache updates.These categories represent different update targets rather than a strict progression in adaptation strength.
  • Input adaptation: Input adaptation modifies test inputs instead of model parameters, preserving the task model when parameter access is unavailable but potentially adding latency or removing task-relevant content.
  • Partial parameter updates: Partial-parameter updates modify a designated subset of weights, aiming to preserve remaining parameters, mitigate forgetting, and keep adaptation costs tractable.Normalization-based and selective-update methods identify subsets through statistics, learned selection, or parameter importance.
  • Auxiliary parameter updates: Auxiliary-parameter methods freeze the pretrained backbone and add prompts, adapters, or low-rank modules that absorb task- or domain-specific shift.This separates pretrained knowledge from test-time adjustments, making adaptation easier to reverse and store.
  • Backbone and full-model updates: Backbone-only updates alter representations while keeping the task head fixed, whereas full-model updates also change the decision boundary and can correct broader errors.Full-model adaptation therefore makes unreliable feedback consequential for both representation and final prediction.
  • External memory or cache updates: External-memory methods update model-external state from deployment observations or feedback and reuse the acquired knowledge in later predictions.Class-prototype methods compress incoming evidence into compact summaries for efficient reuse.

3.4. Temporal Horizons of Test-Time Learning

Temporal design in test-time learning has two independent dimensions: the evidence used for each update and the lifetime of the resulting writable state. Separating these horizons clarifies how methods trade signal reliability, latency, experience reuse, and error accumulation.

  • Supervision horizon: The supervision horizon specifies whether each update uses a single input or token, a batch or history, or a structured trajectory.Single-sample methods use local evidence and derived views, while multi-sample methods exploit relationships across batches, histories, sequences, or interactions.
  • Persistence horizon: The persistence horizon specifies how long writable state—including parameters, statistics, prompts, fast weights, modules, or memory—is retained.Episodic persistence resets state within bounded deployment units; continual persistence carries it across the deployment stream.
  • Joint temporal design: A method may combine local evidence with stream-level retention or trajectory evidence with episodic reset, so supervision and persistence should be selected separately.The same temporal design applies to distribution-shift adaptation, context memorization, personalization, and interactive learning.

3.5. Test-Time Learning for Adaptation

Test-time learning adapts models to deployment distributions through training-prepared or fully test-time objectives, while practical effectiveness depends on feedback quality, access constraints, and stream structure. Theory supports specific mechanisms and assumptions rather than universal guarantees.

  • Distribution shift: Domain shift occurs when the deployment distribution differs from the training distribution, with covariate, label, and compound shifts among the relevant cases.Covariate shift changes inputs while preserving labeling; label shift changes class prevalence while preserving class-conditional inputs.
  • Foundational paradigms: Test-time training prepares an auxiliary objective during source training, whereas fully test-time adaptation constructs objectives from signals available to an off-the-shelf model.TTT depends on auxiliary–main task alignment; FTTA is more practical without source-training changes but is more exposed to unreliable feedback.
  • Test-time objectives: Entropy minimization, consistency, pseudo-labels, and related signals broaden fully test-time adaptation beyond redesigned source-training pipelines, but remain dependent on accurate and diverse self-generated feedback.
  • Deployment constraints: Gradient access and memory overhead constrain deployment: TENT requires 5,165 MB versus 832 MB for inference on ViT-Base in the cited comparison.These constraints motivate forward-only and gradient-free adaptation methods.
  • Theoretical understanding: Theoretical results establish guarantees only under specific objectives, feedback mechanisms, update rules, and deployment streams, not universally for test-time adaptation.
  • Failure and stability: Reliable adaptation requires informative signals, controlled updates, and mechanisms for detecting harmful shifts because surrogate optimization can reverse gains, amplify pseudo-label errors, or cause collapse.Non-stationary-stream analyses additionally expose an adaptivity–information trade-off under assumptions about shift structure and proxy–task alignment.

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

Beyond distribution-shift adaptation, test-time learning can accumulate deployment experience, personalize behavior, specialize models to recurring contexts, and support recursive self-improvement. These approaches range from storing and retrieving information to modifying agents and the processes that improve them.

  • Beyond Adaptation: Beyond-adaptation test-time learning reuses deployment experience to improve future behavior through evolving memory, personalization, specialization, interactive feedback, and recursive self-improvement.These objectives extend beyond merely recovering source-domain performance.
  • Evolving Memory: Test-time learning can store, organize, and retrieve information from observations and feedback so later predictions or actions reuse deployment-acquired knowledge.Design objectives include memorizing long context and accumulating experience for future action refinement.
  • Evolving Memory: TTT-Layer treats a sequence model’s hidden state as a learnable memory updated per token, reducing long-context sequence complexity from O(n^2) to O(n).The update uses a self-supervised reconstruction objective over the current token.
  • Evolving Memory: Evolving-memory systems retain reusable experience across task or interaction streams, including retrievable histories, verified physical principles, and task-specific knowledge across episodes.Examples span language-model agents and embodied agents.
  • Recursive Self-Improvement: Recursive self-improvement forms a loop in which systems propose changes, evaluate them, retain successful modifications, and use modified systems in later cycles.The defining distinction is whether retained modifications change how subsequent improvement is produced.
  • Recursive Self-Improvement: Recursive self-improvement progresses from evolving solutions under fixed loops to modifying agent implementations and then improving the process that generates future modifications.Process-level systems require held-out evaluation, regression testing, version isolation, and rollback to limit objective drift and regression.
  • Personalization: Personalization specializes a generic model to individual users, patients, or instances, addressing individual-specific failures distinct from distribution-shift degradation.Applications include user-preference generation, human-centered recognition, and biomedical modeling.
  • Specialization: Specialization tailors general-purpose models to particular deployment contexts, often using limited environment- or task-specific data for physical environments and robotics.Robotics examples adapt to local geometry, dynamics, and task semantics.

4. Test-Time Scaling

Test-time scaling improves inference by allocating controllable additional computation while keeping pretrained parameters frozen. The section covers longer reasoning, repeated sampling, consensus, search, verification, and adaptive allocation, emphasizing that gains depend on reliable candidate generation and evaluation.

  • Problem Formulation: Test-time scaling allocates additional, controllable inference computation while keeping pretrained model parameters frozen.The compute budget may specify samples, reasoning steps, search depth, verification rounds, or tool calls.
  • Inference-Time Compute: Scaling can extend one reasoning trajectory, generate multiple candidates, or explore structured intermediate-state spaces.Adaptive methods determine how much computation each input or reasoning stage receives.
  • Best-of-N Sampling: Best-of-N sampling generates N candidates and selects one using a scoring criterion, but its value depends on reliable candidate comparison.Repeated sampling can increase the chance that a correct answer appears without ensuring that it is selected.
  • Consensus Mechanisms: Consensus methods select answers through agreement across sampled reasoning paths, yet correlated samples or ambiguous equivalence can produce confident shared errors.Majority-vote performance may improve and then decline as the number of calls grows.
  • Search-Based Inference: Search preserves relations among partial states, enabling lookahead, revision, backtracking, and evaluator-guided expansion rather than selection only after complete generation.Tree and graph methods organize alternatives and can use world models, rewards, reflection, or learned value functions.
  • Theoretical Understanding: Additional computation is conditional rather than monotonic: it helps when the system produces useful candidates and evaluates and allocates them reliably.Theoretical analyses identify risks including reward hacking, verifier dependence, overthinking, and underthinking under fixed budgets.

4.3. Scaling with External Resources

External-resource scaling extends inference beyond the model’s internal computation by supplying evidence, execution, evaluation, or complementary agent capabilities. These resources can improve reasoning and selection, but their benefits depend on relevance, reliable interfaces, informative evaluators, and non-correlated contributions.

  • Resource Motivations: External resources address missing evidence, exact operations, executable actions, or independent candidate assessment that internal computation alone cannot provide.The survey distinguishes retrieval, tools, verifiers, critics, re-rankers, and multi-agent inference.
  • Retrieval: Adaptive retrieval couples evidence acquisition with reasoning through on-demand retrieval, confidence triggers, iterative queries, or structured search.Examples include Self-RAG, FLARE, IRCoT, Search-o1, Search-R1, and dynamic knowledge-graph retrieval.
  • Tool Use: External tools delegate selected operations to executors and can return feedback that informs later decisions, validation, or revision.ReAct interleaves reasoning, actions, and observations, while PoT separates semantic reasoning from program execution.
  • Verification and Selection: Verifiers, critics, and re-rankers improve selection or refinement by evaluating complete outputs or intermediate reasoning states.They can use scalar scores, textual critiques, process-level feedback, aggregated judgments, or multiple verifiers.
  • Multi-Agent Inference: Multi-agent systems gain from complementary information, expertise, tools, or error patterns rather than agent count alone.Correlated errors, premature convergence, and communication or aggregation costs can make additional agents ineffective or harmful.

4.4. What Scaling Improves and Enables

Test-time scaling improves inference by generating, organizing, evaluating, and revising multiple candidate reasoning paths or decisions. Its benefits extend beyond accuracy to reasoning quality, reliability assessment, and control and planning.

  • Reasoning Quality: Additional computation improves reasoning by exploring multiple decoding paths and aggregating or comparing their intermediate steps.Chain-of-Thought increases depth, while self-consistency and repeated sampling increase breadth.
  • Reasoning Quality: Tree- and graph-based search, process supervision, verifiers, and adaptive allocation organize trajectories and concentrate computation where it is useful.These mechanisms structure alternatives, revisit intermediate states, and filter or redirect candidate reasoning.
  • Reliability and Selection: Verification, candidate selection, and confidence assessment help determine whether outputs should be accepted, revised, or withheld.Tools, retrieval, reflection, confidence-aware aggregation, and uncertainty estimation support these distinct roles.
  • Reliability and Selection: More samples or consensus can improve selection without calibrating confidence when reasoning paths share correlated errors.Reliable scaling therefore requires both effective evaluation or aggregation and informative uncertainty signals.
  • Decision Quality in Control and Planning: In control and planning, additional computation organizes, compares, and updates candidate decisions as feedback arrives.Decomposition, hierarchical execution, trajectory evaluation, environment interaction, and world models support replanning, but inaccurate signals can optimize the wrong path.

4.5. Limitations of Pure Scaling

Pure test-time scaling improves current inference episodes but introduces latency, resource costs, limited persistence, saturation, and instability. Its deployment value depends on reliable marginal benefit, appropriate allocation, and whether gains need to persist across queries.

  • Deployment Costs and Scope: Pure scaling adds latency and resource costs, while its gains may remain local to the current episode or become redundant and unstable.These limitations directly affect deployment consequences.
  • Deployment Costs and Scope: Repeated sampling, search, and verification require additional model calls and intermediate-state expansion, scoring, and revision.Tree search can be substantially slower than simpler baselines while yielding limited incremental gains in some settings.
  • Deployment Costs and Scope: Compute-optimal and adaptive allocation reduce over-processing but cannot eliminate generation and evaluation costs, so deployment must jointly consider latency, memory, serving cost, and accuracy.
  • Persistence: Pure scaling usually improves only the current episode because search results do not automatically accumulate for later queries.Cross-query persistence requires writing results into parameters or memory or transferring them through offline learning.
  • Saturation and Instability: More computation can saturate as candidate diversity plateaus, producing redundant outputs and smaller marginal gains.Longer reasoning may also add noise instead of useful progress.
  • Saturation and Instability: Scaling can become unstable when self-critique reinforces incorrect judgments or coordinated agents propagate shared errors.Useful scaling requires informative generation, reliable evaluation, and stopping before marginal gains are exhausted.
  • Overall Trade-off: Deployment value depends on whether reliable marginal benefit exceeds latency and resource cost and whether effects are appropriately episode-local or persistent.

5. The Intersection of Learning and Scaling

Learning and scaling intersect through scaling for learning, learning to scale, and joint systems that update active state during deployment. The survey distinguishes reusable supervision produced by computation from learned control of computation and closed-loop adaptation.

  • Scaling for Learning: Scaling for learning converts additional inference or inference-like computation into signals for model learning.This relationship often occurs during training or data construction and is outside TTI when its outputs become reusable supervision or parametric capability.
  • Scaling for Learning: Multiple generations, consensus, rationale bootstrapping, and filtered samples provide answer- or rationale-level targets when direct supervision is unavailable.Examples include self-consistency, STaR, and ReSTEM.
  • Scaling for Learning: Verification filters, ranks, or annotates self-generated outputs through answer-level, process-level, preference, and self-judgment feedback.Process supervision can derive denser step labels from rollouts or intermediate assessments.
  • Scaling for Learning: Search and interaction generate trajectories with intermediate decisions, branches, recovery paths, and supervision for policies or values.Failure-recovery data can become contrastive or corrective supervision in long-horizon settings.
  • Scaling for Learning: Search-gain distillation compresses expensive candidate selection or search into parameters, reducing future inference cost while preserving much of the improvement.
  • Learning to Scale: Learning to scale uses learned controllers to allocate computation, route among inference mechanisms, or prepare procedures that exploit scaling effectively.A fixed controller supports TTS deployment, whereas updating it during deployment constitutes TTL.
  • Learning to Scale: Adaptive allocation matches tokens, samples, or search effort to expected utility rather than assigning every input the same budget.This avoids over-processing easy instances while retaining computation for difficult ones.
  • Learning to Scale: Controllers can route between direct answers, reasoning, planning, retrieval, tools, search, or heterogeneous models.Routing may occur upfront or through cascades that inspect earlier outputs before escalation.

6. Applications

The survey applies test-time intelligence across vision, video, generative models, 3D, spatial reasoning, speech, and agentic AI. These applications use adaptation, verification, memory, interaction, and additional computation to address changing observations and specialized deployment conditions.

  • Vision: Vision is a mature TTI setting because sensor, weather, acquisition, and temporal shifts alter observations while labels are usually unavailable.Conventional TTL methods address image classification, segmentation, and detection.
  • Vision: Video adaptation constructs signals across ordered frames or temporal segments, including spatio-temporal alignment, self-supervision, and synchronized prompt tuning.
  • Generative Models: Generative-model TTI refines restoration, synthesis, video consistency, 3D representations, and stochastic generation trajectories through extra inference computation.Instance-internal methods also optimize directly on observations or image-specific self-supervised pairs.
  • Generative Models: Video-generation methods target motion and appearance customization, pose transfer, long-video stabilization, and compositional alignment.
  • 3D: Test-time 3D methods either optimize instance-specific representations from pretrained generative priors or update models and representations from incomplete observations.The routes include NeRF distillation, mesh refinement, shape completion, point-cloud correction, and test-time training.
  • Generative Models: Diffusion TTI searches over stochastic denoising trajectories by comparing initial noises, branching or resampling intermediate states, and evaluating partial generations.
  • Spatial Reasoning: Spatial reasoning uses verifier–refiner loops, imagined viewpoints, multi-view evidence, and adaptive visual imagination to improve geometric predictions.Explicit test-time updates remain under-explored in this area.
  • Speech and Audio: Speech TTA addresses acoustic shifts from speaker variability, background noise, and corpus mismatches, but sequential audio dynamics limit general-purpose methods.

7. Open Challenges and Future Directions

The survey identifies reliability, efficiency, evaluation, realism, theory, and security as central challenges for deploying TTI at scale, while outlining directions that extend TTI to memory, research, science, diffusion, and human interaction.

  • Core challenges: Long-horizon deployment is difficult because arbitrary open-world data makes noisy unsupervised signals for test-time learning less reliable.
  • Core challenges: Reliable objectives remain essential because learning-based methods need unsupervised update signals, whereas scaling-based methods need criteria to assess, verify, or rank outputs.
  • Core challenges: Inference-time improvement introduces efficiency costs, motivating unreliable-sample filtering, learned optimizers, forward-only methods, and applications to resource-constrained settings.
  • Core challenges: Current TTI benchmarks often use synthetic corruptions or benchmark-specific reasoning tasks, motivating evaluations with mixed, evolving, multimodal, personalized, temporal, and interactive shifts.
  • Core challenges: Existing theory analyzes isolated mechanisms under specific assumptions but lacks a unified account of feedback-driven test-time intelligence.
  • Safety and future directions: Scalable long-term memory is a promising direction, building on test-time training methods that memorize historical context and improve memory capacity, reliability, transfer, and modality coverage.
  • Safety and future directions: Adversarial inputs can poison adaptive states or manipulate scaling evaluators, so TTI needs standardized attack protocols and defenses for unreliable feedback, harmful state changes, and recovery.
  • Safety and future directions: TTI could support scientific discovery by using simulators, experiments, constraints, domain knowledge, and tools to explore large solution spaces and adapt inference to evolving environments.

8. Conclusions

The survey presents TTI as a unified framework for deployment-time improvement through state updates and additional inference-time computation. It synthesizes methods, applications, and remaining challenges toward adaptive and self-improving AI systems.

  • TTI connects test-time learning, test-time adaptation, and test-time scaling by organizing methods around state updates and inference-time computation.
  • The survey reviews representative methods and applications across vision, language, multimodal learning, generative models, robotics, and healthcare.
  • Key challenges include reliable objectives, long-horizon stability, efficiency, realistic evaluation, theoretical foundations, and safety risks.
Loading 2609.01679v1…