Think Fast, Act Faster: How 'Thinking-by-Doing' Is Rewiring LLM World Models

Opening — Why this matters now

Agentic AI is finally tipping from novelty to necessity. Models are no longer asked to answer questions — they’re asked to navigate, plan, act, and recover from their own mistakes. But here’s the uncomfortable truth: most LLMs still think like academics when the world increasingly demands engineers. They simulate long chains of imaginary transitions, hallucinating entire environments inside their heads. It’s elegant — and disastrously brittle.

The paper Thinking by Doing fileciteturn0file0 surfaces an unavoidable shift: agents that reason by interacting, not by daydreaming. And this shift is more than a training trick — it’s a governance problem, a capability accelerator, and a coming expectation for enterprise automation.

Background — The rise and limits of monolithic reasoning

Long-context chain-of-thought and massive internal simulation have powered most reasoning-focused LLMs. The model thinks through dozens of hypotheses before giving a final answer. For math and code, this works. For physical or environmental tasks, it breaks — spectacularly.

On page 2, the paper illustrates how monolithic reasoning forces the model to simulate full navigation paths, reinforcing incorrect assumptions and drifting further from ground truth. This is reasoning as wishful thinking.

Multi-turn interaction flips the relationship: instead of reasoning before doing, the model reasons because it is doing. Environmental feedback isn’t a final check — it’s the substrate from which knowledge is grown.

This reflects a broader shift: structured, human-designed cognitive patterns are reaching diminishing returns. The old recipe — perception, planning, prediction — is rigid. Enterprise AI systems need agents that learn procedures through action, not annotation.

Analysis — What WMAct actually does

WMAct (World-Model internalization through Active interaction) is built around two deceptively simple mechanisms:

1. Reward Rescaling: penalize useless actions

On page 5, the authors show a problem as old as reinforcement learning itself: brute-force exploration. Even powerful models spam irrelevant moves because the environment doesn’t punish inefficiency.

WMAct scales rewards by the share of effective actions:

$$R_{scaled} = R_{outcome} \times \frac{N_{eff}}{N}$$

This is subtle governance by incentive design. The model stops acting like a child mashing every button on a controller and starts behaving like an operator who knows actions have cost.

2. Interaction Frequency Annealing: force internalization

The central danger of interactive agents is over-reliance on feedback. If you let the model query the environment forever, it never learns to think.

The annealing mechanism (formula on page 5) gradually reduces the allowed number of interaction turns:

$$L_{max} = \frac{\bar{L} + L’_{max}}{2}$$

Early training: explore freely. Late training: compress the world model.

This is cognitive discipline enforced at the training-loop level.

Why this combination works

WMAct doesn’t impose human-designed reasoning steps. Instead, it shapes the pressure under which the model learns. The agent starts with open-ended exploration, then is squeezed into forming an internal, portable world model.

In other words: WMAct trains agents to get smarter by having fewer chances to be stupid.

Findings — Evidence that reasoning is becoming embodied

The empirical results across Maze, Sokoban, and Taxi (Table 1, page 6) show:

• Single-turn accuracy approaches multi-turn accuracy (Figure 4).
• Generalization improves even on harder maps (Hard-1 and Hard-2).
• Sokoban-trained WMAct boosts math and logic benchmarks (Table 2).

A particularly revealing curve:

Single-turn performance grows to match multi-turn

This means the agent is no longer dependent on interaction. It has internalized the environment.

A simple visualization

The visual traces in Section C (pages 14–15) reinforce the point: WMAct agents display reflection, adaptation, and recovery — behaviors not explicitly trained but emergent under pressure.

Implications — What this means for real-world AI automation

Industrial and enterprise automation is shifting from static workflows to dynamic agents. WMAct’s findings point toward several tangible transformations.

1. Agent governance will move from rules to incentives

As this paper shows, behavior can be shaped without prescribing cognitive steps. This offers a new toolkit for controlling autonomous systems: shape the reward landscape, not the thought structure.

2. Corporate AI agents will need embodied reasoning

Customer support bots, logistics planners, and workflow orchestrators all operate in environments with real feedback loops — databases, APIs, human inputs. Embodied reasoning increases reliability because the agent becomes less sensitive to brittle internal simulations.

3. Fewer steps, lower cost, faster throughput

Single-turn reasoning that carries world-model knowledge reduces inference cost. Companies deploying agent swarms will feel this directly on the GPU bill.

4. The capability frontier is shifting

Agents that learn through action will outperform models that reason in isolation. WMAct’s superior transfer performance on MMLU-Pro, GPQA-Diamond, and LiveBench (Table 2, page 7) demonstrates that embedded interaction training strengthens general reasoning — not just task-specific heuristics.

Conclusion — Where this is heading

“Thinking by doing” isn’t a slogan — it’s a necessary re-alignment for agentic AI. WMAct underscores a deeper truth: models must learn to compress experience, not merely recite logic. The next generation of enterprise agents will be shaped by incentives and constraints rather than scripted cognitive routines.

And for those building AI automation platforms, the lesson is clear: your agents need environments, feedback, and disciplined interaction — not more prompts.

Cognaptus: Automate the Present, Incubate the Future.