PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative AI Act II: Test Time Scaling Drives Cognition Engineering

Shijie Xia, Yiwei Qin, Xuefeng Li, Yan Ma, Run-Ze Fan, Steffi Chern, Haoyang Zou, Fan Zhou, Xiangkun Hu, Jiahe Jin, Yanheng He, Yixin Ye, Yixiu Liu, Pengfei Liu

arXiv:2504.13828v3cs.CLcs.AI

TL;DR

Existing LLM development largely scaled knowledge acquisition, while cognition engineering addresses the need for deeper reasoning and thinking capabilities. The paper defines this paradigm, surveys test-time scaling, training, applications, infrastructure, and tutorials, and reports substantial performance improvements across domains. It also notes limitations in reward modeling and iterative self-reinforced learning as task complexity and training iterations increase.

  • Problem

    Existing LLMs primarily organize and retrieve knowledge, motivating a paradigm focused on deeper reasoning, deliberate thinking, and cognitive capability development.

  • Method

    The paper defines cognition engineering and systematically surveys test-time scaling, specialized training strategies, applications, infrastructure, future directions, and implementation tutorials.

  • Results

    The paper reports substantial improvements in model performance across domains including mathematics and multimodal understanding.

  • Takeaways & Limitations

    Cognition engineering marks a second act of generative AI centered on systematically developing thinking capabilities through test-time scaling and training.

  • Takeaways & Limitations

    Model-based rewards can cause reward hacking through out-of-distribution queries and reward mis-specification, while iterative self-reinforced learning may plateau or decline after a few iterations.

Abstract

from arXiv · show

The first generation of Large Language Models - what might be called "Act I" of generative AI (2020-2023) - achieved remarkable success through massive parameter and data scaling, yet exhibited fundamental limitations such as knowledge latency, shallow reasoning, and constrained cognitive processes. During this era, prompt engineering emerged as our primary interface with AI, enabling dialogue-level communication through natural language. We now witness the emergence of "Act II" (2024-present), where models are transitioning from knowledge-retrieval systems (in latent space) to thought-construction engines through test-time scaling techniques. This new paradigm establishes a mind-level connection with AI through language-based thoughts. In this paper, we clarify the conceptual foundations of cognition engineering and explain why this moment is critical for its development. We systematically break down these advanced approaches through comprehensive tutorials and optimized implementations, democratizing access to cognition engineering and enabling every practitioner to participate in AI's second act. We provide a regularly updated collection of papers on test-time scaling in the GitHub Repository: https://github.com/GAIR-NLP/cognition-engineering

Three Scaling Phases

The three scaling phases progressively transform model knowledge from isolated concepts to denser associations and finally to dynamically constructed reasoning pathways. Test-time scaling is presented as the phase that enables cross-domain inference and advanced scientific reasoning through extended computation.

  • Stage 1: Pre-training Scaling: Pre-training scaling forms specialized knowledge islands by increasing model size and training data, but connections among concepts remain limited.The initial representation primarily captures basic knowledge and innate relationships.
  • Stage 2: Post-training Scaling: Post-training scaling densifies knowledge islands with learned connections, especially among closely related concepts, enabling more sophisticated associations.These connections improve integration but do not yet provide the broad bridges associated with test-time reasoning.
  • Stage 3: Test-time Scaling: Test-time scaling enables complex cross-domain reasoning by forming dynamic pathways between previously disconnected concepts through extended computation.It can connect principles such as Universal Gravitation to applications including Orbital Motion or Falling Objects.
  • Conclusion: The progression reframes AI advancement as developing cognitive capabilities, not merely accumulating parameters or data.The paper compares test-time deliberation with sustained thought in human expert problem solving.

The Practitioner’s Roadmap: How to Apply Test-Time Scaling to your Applications?

The section introduces a workflow for applying test-time scaling in a specific domain, while directing readers to the main paper for details.

  • Workflow: Figure 2 presents a workflow for applying test-time scaling in a specific domain.The caption provides no further procedural details.

1 Introduction

The paper frames cognition engineering as generative AI’s second act, shifting LLMs from knowledge management and retrieval toward deliberately engineered thinking capabilities. It defines the paradigm, motivates its timing, and surveys technical foundations, applications, infrastructure, tutorials, and future directions.

  • Background: Earlier LLMs organized accumulated human knowledge through extensive pretraining and fine-tuning, supporting conversation, information retrieval, and content generation.This background establishes the knowledge-management role that cognition engineering seeks to extend.
  • Definition: Cognition engineering systematically develops AI thinking capabilities through test-time scaling, human cognitive pattern distillation, and AI-driven discovery.The definition explicitly distinguishes this approach from traditional pretraining-centered development.
  • Deep Cognition: Cognition engineering focuses on deep cognition involving complex reasoning, deliberate thinking, connections among disparate concepts, and novel insights.Its target includes understanding why and how, not only what.
  • Paradigm Shift: Test-time scaling marks a paradigm shift from knowledge-retrieval systems toward deeper searches and thought construction during inference.The paper presents this shift as generative AI’s second act.
  • Roadmap: The paper will clarify cognition engineering’s concepts, explain its critical timing, analyze test-time scaling and training strategies, and examine applications and broader implications.It also covers infrastructure, future directions, and practical implementation tutorials with code examples.

2 What – Cognition Engineering Definition

Cognition engineering is defined as an engineered transition from knowledge toward wisdom-level AI cognition within the DIKW framework. It emphasizes deep reasoning, metacognition, creative problem solving, and dynamic thought processes rather than static knowledge imitation.

  • 2.1 Cognition: The DIKW framework models cognition as a hierarchy from raw data through contextualized information and applicable knowledge to profound wisdom.Cognition engineering is positioned as the step toward wisdom-level capability.
  • 2.1 Cognition: Wisdom-level cognition involves understanding deep principles, creative thinking, judgment, multi-layered reasoning, metacognition, adaptability, and conceptual abstraction.It concerns knowing why and how in addition to knowing what.
  • 2.2 Engineering: Engineering denotes intentional construction through targeted interventions in training methodologies and inference optimization rather than relying solely on emergence.This makes the paradigm a designed process for moving systems upward in the DIKW hierarchy.
  • 2.3 Cognition Engineering: Cognition engineering constructs and optimizes AI systems to ascend from knowledge to wisdom through design patterns, training strategies, and computational allocations.The resulting systems are intended to exhibit deep thinking, complex reasoning, and creative problem-solving abilities.
  • Methodological Distinction: Compared with traditional models that imitate human outputs, cognition engineering emphasizes imitation and development of human thought processes.The distinction targets wisdom-level cognitive characteristics rather than behavior alone.
  • Methodological Distinction: Cognition engineering also emphasizes dynamic thinking and knowledge creation, contrasting with fixed capabilities and retrieval of existing knowledge.These contrasts are presented as further differences from traditional LLM development.

3 Why & Why Now – Technical Foundation

Cognition engineering responds to LLM limitations in complex reasoning, knowledge updating, and wisdom-level applications, enabled by richer knowledge foundations, test-time scaling, and self-training.

  • Why: LLMs struggle with multi-step reasoning, knowledge updating and creation, and the wisdom-level capabilities demanded by complex applications.These limitations include difficult mathematical proofs, scientific problem-solving, multidimensional analysis, adapting to new developments, and generating original hypotheses or conclusions.
  • Why Now: Cognition engineering emerges as technological breakthroughs create conditions for moving from knowledge management toward deeper cognitive capabilities.The paper identifies three enabling pillars: transformed knowledge acquisition, test-time scaling, and advanced self-training.
  • Knowledge Foundation: Richer, curated training corpora provide the extensive embedded knowledge required for models to perform deeper thinking.These corpora combine scientific and technical materials, mathematics, programming code, and specialized-domain knowledge.
  • Test-Time Scaling Foundation: Test-time scaling reallocates inference computation through stepwise reasoning, tree search, self-correction, and verification to create a cognitive workspace.These methods extend beyond fixed-length, single-pass generation and support exploration, evaluation, error identification, and refinement.
  • Self-Training: Advanced self-training offers a route to eliciting cognitive capabilities beyond what can be obtained from expert human cognition data alone.The paper links reinforcement learning with verifiable rewards to opportunities for AI self-discovery and superhuman performance.
  • From Theory to Practice: Test-time scaling is presented as the most immediate practical bridge from cognition engineering’s theoretical foundations to real-world implementation.The following methods examine how inference-time computation can extend and deepen reasoning processes.

4 How – Part I: Test-Time Scaling Methods

Test-time scaling methods allocate inference computation through search, selection, verification, and iterative reasoning, with efficiency determined by computation budgets and method design.

  • Framework: Test-time scaling abstracts inference as a search strategy M guiding generator g, with scaling dimensions determining computation cost and performance.Efficiency maps computation budget C and method M to performance, while strategies improve through optimizing individual methods or combining methods.
  • Methods Overview: The framework covers parallel sampling, tree search, multi-turn correction, and long CoT, comparing their scaling dimensions, efficiency, and combinations.The paper analyzes construction, scaling laws, individual optimization, cross-method comparisons, and combined strategies.
  • Parallel Sampling: Parallel sampling generates N independent candidate responses and selects among them using Best-of-N scoring, majority voting, or a combined strategy.Best-of-N depends on a scoring function, majority voting selects the most frequent answer, and combined methods trade scoring targeting against voting robustness.
  • Parallel Sampling: External verifiers such as code interpreters and math proof checkers can score candidate responses, while trained outcome or process reward models provide alternatives.Process reward models evaluate intermediate steps and aggregate them into an overall response score.
  • Scaling Laws: Pass@N grows steadily and often log-linearly with the sampling number N across models and tasks.Pass@N measures the probability that at least one of N candidates is correct, while Maj@1 and BoN measure practical selection performance.
  • Scaling Laws: Increasing Pass@N does not guarantee real-world performance because selecting correct responses requires reliable verifiers, which are unavailable or hackable for many tasks.False positives can arise when code passes limited tests or mathematical solutions reach correct answers through incorrect reasoning.
  • Scaling Laws: With a fixed inference budget, larger models with fewer samples can outperform smaller models with more samples on code, whereas smaller models can be more effective for mathematics.The preferred allocation depends on task and model size rather than sampling count alone.
  • Parallel Sampling: Selection precision improves when methods use stronger evaluators, weighted voting, or multiple verifiers.PRM outperforms ORM in BoN settings, weighted majority voting improves over majority voting or BoN at large sampling numbers, and MAV uses multiple verifiers.

4.2 Tree Search

Tree search frames test-time reasoning as guided exploration over structured problem-solving spaces, with node granularity, search algorithms, and value functions determining efficiency and deliberation. Scaling search breadth and depth can improve performance, but gains eventually saturate and depend on reliable evaluation.

  • Tree Search Framework: Tree search explores a structured space of problem-solving approaches, using algorithms and value functions to guide deliberate planning.The framework searches nodes in space S and evaluates them for selection, pruning, or exploitation.
  • Search Space: Search granularity ranges from tokens to steps and complete solutions, trading optimality and computational cost against efficiency.Token-level search is fine-grained and costly, step-level search balances granularity and efficiency, and solution-level search updates whole responses through critique or revision.
  • Value Function: Value functions estimate candidate-node quality through self-evaluation, specialized evaluators, action likelihoods, self-consistency, or rollout-based estimates.LATS combines language-model and self-consistency scores, while rollout values can be updated from terminal-state rewards.
  • Search Algorithm: Tree-search algorithms include breadth-first methods such as beam search and A*, depth-first backtracking, and Monte Carlo Tree Search with selection, expansion, evaluation, and backup.These algorithms differ in how they explore promising nodes and alternatives.
  • Scaling Laws: Performance improves with broader or deeper search, but eventually saturates as candidate diversity declines; larger process-reward models can further improve results.Scaling includes more MCTS rollouts, larger beam sizes, and greater A* step limits, while improved value-function reliability remains important.
  • Scaling Efficiency: For code generation, MCTS can outperform beam search at equal computation time, while A* can be more time-efficient than MCTS and alternative methods.The most suitable algorithm depends on task characteristics and computational constraints.

4.3 Multi-turn Correction

Multi-turn correction improves responses through iterative generation, feedback, and refinement. Its benefits depend strongly on feedback reliability and correction ability: revision scaling helps with reliable correctors but can hurt intrinsic self-correction without additional training.

  • Framework: Multi-turn correction iteratively generates an answer, obtains feedback, and refines the response until a stopping condition is met.Feedback may be omitted when the system directly refines the initial response.
  • Feedback and Refinement: Feedback can be self-generated or supplied externally by tools, knowledge sources, or oracle labels, while refinement can use the same or a specialized model.Self-feedback includes reflective prompting and execution traces; external feedback includes interpreters, proof checkers, simulators, and ground-truth answers.
  • Scaling Laws: Reliable external feedback enables performance gains from additional revision steps until saturation, whereas intrinsic self-correction can worsen as revisions increase when critique ability is weak.This contrast makes feedback quality and refinement capability central constraints on scaling.
  • Scaling Laws: Additional training can improve refinement ability, after which performance may continue growing with more revision steps before eventually saturating.Snell et al. (2024) report gains beyond the revision count used during training after fine-tuning correction ability.
  • Scaling Efficiency: Improving scaling efficiency therefore requires higher-quality feedback mechanisms or stronger refinement models, including reference-free evaluation, task decomposition, confidence estimation, and targeted training.These strategies address the reliability and capability bottlenecks identified for multi-turn correction.
  • Long Chain-of-Thought: Long chain-of-thought reasoning supports reflection, backtracking, verification, divergent thinking, and internal thinking, but can overthink simple problems and allocate computation uniformly across tokens.The overthinking problem motivates more selective allocation of test-time computation.

4.5 Comparisons of Test-Time Scaling Methods

Test-time scaling methods differ in performance, cognitive behavior, adaptivity, controllability, simplicity, training requirements, and compatibility. Long chain-of-thought has the strongest overall profile but requires additional training, while combining methods can exploit complementary strengths.

  • Performance: Long chain-of-thought generally offers the highest performance ceiling, especially on difficult problems, while beam search can excel under limited budgets and best-of-N sampling on simpler tasks with more compute.These complementary performance regimes make absolute rankings difficult across implementations and budgets.
  • Cognitive Behaviors: Long chain-of-thought exhibits the most integrated cognitive behaviors, including reflection, backtracking, and divergent thinking, whereas tree search and correction rely on predefined external structures.External structures can limit adaptation to specific problems.
  • Adaptivity: Adaptivity depends on stopping conditions: verifier-based termination can allocate more computation to harder problems, while fixed hyperparameters provide no additional adaptivity.Parallel sampling normally assigns identical sampling counts across queries.
  • Controllability and Simplicity: Long chain-of-thought is simpler to implement and has stronger overall controllability and adaptivity, but unlike most other methods it requires additional training to elicit.Other methods commonly require evaluators or multiple components, increasing deployment costs.
  • Compatibility and Ensembles: All test-time scaling methods can be combined, with parallel sampling especially easy to integrate; ensembles can therefore allocate computation across complementary search and correction behaviors.The paper presents combinations involving tree search, multi-turn correction, and long chain-of-thought.
  • Ensemble Methods: Accounting for overthinking in voting can outperform standard majority voting or a single expensive response, while shorter parallel responses can be more compute-optimal than forcing longer chain-of-thought.High overthinking correlates with lower performance in some math and agentic tasks.

5 How – Part II: Training Strategies for Test-Time Scaling

The paper presents reinforcement learning and supervised fine-tuning strategies for eliciting long chain-of-thought reasoning, while emphasizing reward design, data, model choice, and iterative self-improvement as key training considerations.

  • 5.1 Reinforcement Learning: Long-CoT reinforcement learning improves reasoning on mathematics and code while eliciting self-reflection and self-correction, but depends on robust reward signals.Rule-based rewards support autonomous learning of long-CoT methods, whereas reliable rewards for open scientific questions remain an open challenge.
  • 5.1.2 Reward Function: Model-based rewards generalize across tasks but can cause reward hacking because proxy rewards suffer from out-of-distribution samples and reward mis-specification.During reward overoptimization, proxy scores can rise while golden rewards saturate and decline.
  • 5.1.2 Reward Function: Process rewards can improve reinforcement learning over outcome-only rewards, but require costly data construction and remain vulnerable to hacking.Their gains are less pronounced in reinforcement-learning settings than in parallel response selection.
  • 5.1.3 Policy Model Selection: Policy-model selection matters because pretrained cognitive behaviors and base-model capabilities shape reinforcement-learning outcomes, while RL may not increase large-K Pass@K.Qwen2.5 improves more than Llama-3.2 under identical training, and priming Llama with cognitive traces narrows the difference.
  • 5.1.4 Training Data Construction: RL scaling benefits from larger models, more curated prompts, and curriculum sampling, whereas training-data scaling for supervised fine-tuning often yields limited gains relative to cost.Open-Reasoner-Zero reports continued accuracy and response-length scaling from 57K prompts, while smaller datasets plateau.
  • 5.2 Supervised Fine-Tuning and 5.3 Iterative Self-reinforced Learning: Fine-tuning choices remain unsettled: LoRA can approach full-parameter performance, small models may favor shorter reasoning chains, and SFT depends on teachers and may memorize patterns.The paper also describes iterative self-reinforced learning, in which generated trajectories train successive policy versions, but improvement can plateau or slightly decline after several iterations.

6 How’s Progress – Application So Far

Test-time scaling has expanded across mathematics, coding, multimodal understanding, and embodied AI, using longer reasoning, search, correction, and interaction to improve difficult tasks while leaving important deployment challenges unresolved.

  • Applications: Test-time scaling methods include parallel sampling, tree search, multi-turn correction, and long chain-of-thought reasoning across application domains.The paper organizes these methods alongside SFT, DPO, reinforcement learning, and inference-only strategies.
  • Mathematics: Mathematical reasoning research combines natural-language tasks with verifiable answers and formal-language systems such as Lean and Isabelle.Natural-language reasoning offers more training data, whereas formal reasoning improves process verifiability but has less available data.
  • Coding: Coding benefits from execution verification and self-reflection, but sandboxing, overthinking, and infrastructure remain practical challenges.The paper also highlights gold-medal performance by o1 and o3 variants at the 2024 International Olympiad in Informatics.
  • Embodied AI: Embodied AI uses reasoning, feedback, verification, backtracking, and reflection to improve per-step decisions and high-level planning.Existing systems still lack viable pure-RL approaches and often separate high-level planning from low-level policy execution.
  • Embodied AI: 26.6% accuracy on Humanity’s Last Exam was achieved by Deep Research, surpassing previous state-of-the-art models through scaled reasoning in decision making.The approach synthesizes online sources into comprehensive reports and uses an optimized OpenAI o3 variant.
  • Safety: Strategically increasing inference-time interaction and reasoning steps offers a scalable path toward safer and more robust behavior without retraining.This conclusion is presented across the surveyed applications rather than as a result from one benchmark.

7 So What? – From Scaling to Cognitive Intelligence

Cognition engineering shifts AI development from storing products of human thought toward constructing and using thought processes. It relies on cognitive data, redesigned environments and rewards, and human-AI collaboration to expand reasoning and scientific work.

  • Conceptual shift: Cognition engineering reframes valuable training data as thought processes rather than only the products of human thinking.This transition introduces cognitive data engineering as a distinct development focus.
  • Cognitive data: Cognitive data can come from human artifacts, AI-generated trajectories, and human-AI collaborative generation.These sources combine recorded expertise, self-discovered strategies, machine diversity, and human judgment.
  • Cognitive data: Cognitive data repositories may become strategically important for AI advancement alongside large-scale computing resources.The paper argues that such repositories could drive capabilities beyond natural data collection or synthetic generation alone.
  • Environments and rewards: Reward verification becomes harder as tasks move from objectively checkable mathematics toward research, scientific discovery, and literary creation.These frontier tasks involve planning, creativity, originality, and contextual judgments that resist complete objectification.
  • Human-AI collaboration: Human-AI partnerships support bidirectional strategy exchange, cognitive amplification, shared workspaces, and adaptive communication.AI can manage detail and explore alternatives while humans provide expertise, values, and contextual understanding.
  • Scientific research: Cognition engineering may accelerate scientific research by reducing bottlenecks in hypothesis generation, literature integration, experimental design, and theory refinement.The paper connects this acceleration with broader participation in scientific discovery and large-scale problem solving.

8 Infrastructure

The infrastructure section presents reinforcement learning and Monte Carlo tree search as core technologies for cognition engineering, while emphasizing rollout efficiency, memory management, scheduling, and scaling limitations.

  • Reinforcement learning: PPO-based reinforcement learning alternates between response rollout and model updates that compute rewards, values, advantages, and losses.Dedicated inference backends such as vLLM or SGLang accelerate rollout generation.
  • Reinforcement learning: OpenRLHF and veRL provide open-source support for reinforcement-learning workflows involving multiple models and resource-allocation strategies.The comparison is summarized in Table 10.
  • Scaling challenges: Scaling reinforcement learning to large models remains difficult because long chain-of-thought training can require prohibitive GPU resources and suffer low utilization.A reported 32B-parameter scaling attempt encountered these hardware constraints.
  • Monte Carlo tree search: Large-scale MCTS deployment is limited by insufficient hardware- and software-aware acceleration in existing code repositories.The paper identifies optimization of tree-search infrastructure as an open implementation need.
  • Acceleration: Speculative decoding, shared tree-structured KV caches, and adaptive parallel processing target rollout speed, memory bottlenecks, and path-switching overhead.These techniques address distinct computational constraints in inference and tree expansion.

9 Tutorial

The tutorial provides an end-to-end recipe for reinforcement-learning-based long chain-of-thought training, covering framework selection, models, datasets, GRPO configuration, reward design, and implementation flow. The reported outcome is improved mathematical accuracy and longer solution processes.

  • Setup: The tutorial uses veRL, recommends Qwen2.5 base models, and adopts GRPO as the primary algorithm.Base models are preferred because they retain exploratory capabilities, while the selected datasets are designed to support stable learning.
  • Implementation: The training workflow includes dataset preparation, rollout generation, reward computation, advantage estimation, and policy-loss optimization.The code walkthrough follows these stages through veRL’s trainer and data-processing components.
  • Reward design: The prompt template asks the model to reason before returning an answer inside a boxed tag for automatic evaluation.The reward function assigns 1 point to correct answers and -1 point to incorrect answers.
  • 9.4 Results: Accuracy increases markedly in the tutorial’s mathematical training experiment, as illustrated in Figure 19.The result is presented together with longer generated solution processes.
  • 9.4 Results: Response length increases, producing more detailed solution processes after reinforcement-learning training.The paper reports this result alongside the accuracy improvement in Figure 19.

10 Future Directions

Future cognition engineering research must address architectural efficiency, richer training signals, rigorous reinforcement-learning methodology, and evaluations that measure cognitive processes beyond task accuracy.

  • New architecture: Architectural advances are needed because Transformer limitations constrain efficient test-time scaling for long-context generation.Proposed directions include state space models, linear Transformers, and language diffusion models, alongside supporting theory, infrastructure, and pretrained models.
  • Pretraining on latent thought: Pretraining on latent thought could enrich models beyond datasets containing only human-written text.Recent results motivate incorporating hidden thinking processes and cognitive behaviors into pretraining data.
  • RL scaling: RL scaling remains early-stage, with limited reproducibility, difficult hyperparameter interactions, and insufficient coverage beyond verifiable mathematics and code tasks.Broader applications require investigation of reward hacking, reward reliability, larger-scale accessible experiments, and stronger theoretical analysis.
  • Evaluation: Evaluation should measure reasoning quality, efficiency, and generalizability rather than relying on benchmark scores, word matching, or isolated case studies.Suggested dimensions include reasoning depth, backtracking efficiency, verification quality, metacognitive awareness, and adaptive evaluation protocols.
  • Scientific discovery: Test-time scaling may support scientific discovery by connecting disparate knowledge domains through extended deliberation.Future applications include hypothesis generation, experimental design, and theory formation.

11 Comparison to Existing Work

The paper combines a test-time-scaling-centered survey with a position paper proposing cognition engineering as a unifying framework for developing deep cognitive capabilities in generative AI.

  • Survey perspective: The survey organizes prior work around test-time scaling rather than only system-2 reasoning or long chain-of-thought, covering scaling laws, efficiency, methods, comparisons, and applications.It also provides detailed discussions of reinforcement learning and supervised fine-tuning for long chain-of-thought techniques.
  • Comparison table: Table 11 compares existing surveys across test-time scaling, scaling laws, scaling efficiency, and related topics, with △ marking limited discussion.TTS denotes test-time scaling, while Law and Efficiency denote scaling laws and scaling efficiency.
  • Survey perspective: The paper covers broader test-time-scaling applications and provides future directions for each than previous surveys.Its comparisons and application discussions are presented as more concrete and specific than earlier treatments.
  • Position-paper perspective: The position paper proposes cognition engineering as a framework unifying technologies that enhance model cognitive abilities.It identifies pretraining on latent thought as one potential future direction and distinguishes this conceptual contribution from survey coverage alone.

12 Conclusion

The conclusion frames cognition engineering as generative AI’s second act: a shift from accumulating knowledge toward systematically developing thinking capabilities through test-time scaling and specialized training.

  • 12 Conclusion: Cognition engineering shifts AI development from knowledge accumulation to systematic development of thinking capabilities.The paper associates this shift with deep thinking, complex reasoning, and creative problem-solving across domains including mathematics and multimodal understanding.
Loading 2504.13828v3…