


4.1 System Architecture 4.1.1 General-Purpose AI





4.1 System Architecture 4.1.2 Language Models as Components




4.1 System Architecture 4.1.3 Agents and System Architecture









4.2 Perception and Generation 4.2.1 Multimodality











4.2 Perception and Generation 4.2.2 Modal Representation

























4.2 Perception and Generation 4.2.3 Multi-Modal Fusion





Type | Description | Example |
Early Fusion | Concatenate raw inputs or early features | Multimodal embeddings |
Late Fusion | Process each modality independently, fuse at output | Ensemble or late classifier |
Middle/Intermediate Fusion | Use attention layers to allow one modality to condition on the other | LXMERT, BLIP |




























4.2 Perception and Generation 4.2.4 Modal Generation































4.2 Perception and Generation 4.2.5 Case Study: Nano Banana


















4.3 Memory and Knowledge 4.3.1 Memory










4.3 Memory and Knowledge 4.3.2 External Knowledge and Retrieval























4.4 Action and Execution 4.4.1 Tool Calling


















4.4 Action and Execution 4.4.2 Coding














More thinking changes a candidate. An action may reveal new evidence or change the environment.
Updating context, beliefs or memory is not necessarily a parameter update.
Stop on a checked result, a justified handoff, a budget limit, or lack of a useful next action.
| Current gap | Useful next operation | What it can resolve |
|---|---|---|
| Policy translated incorrectly | Re-read policy; repair formalization | Semantic fidelity |
| Proof search unfinished | Allocate solver time | Formal consequence |
| Filing record unreliable | Query authoritative filing system | Evidence about F |
| Source conflict unresolved | Inspect provenance / request review | Dependence and factual dispute |
P(R ∣ O) = 11/21 concerns review priority. It does not authorize approval.
Choose an operation that addresses the actual uncertainty.



4.5 Planning and Control 4.5.2 System I / II in Agents

| Dimension | Direct response | Additional deliberation |
|---|---|---|
| Computation | Generate the answer directly | Reason, branch, revise or check |
| Useful when | Current policy and evidence suffice | Extra work can resolve uncertainty |
| Failure | Confident shortcut | Repeated error; unnecessary work |
| Measure | Success, actual compute, latency, evidence gained | |
Fast is not automatically shallow. Slow is not automatically correct.
These labels do not establish human-like mental systems inside an LLM.
No fixed “95% / 5%” allocation is assumed. A direct answer still requires neural computation.
| Layer | What changes? | Example |
|---|---|---|
| Training | Model parameters | SFT, RLVR, distillation |
| Decoding / inference | Candidate trajectory and budget | Sampling, search, reasoning tokens |
| Agent control | Next operation and environment state | Call a tool, observe, retry, stop |
A single model can support several modes. A search can repeatedly call a direct policy.
Nucleus sampling is a decoding choice, not a definition of System I.
Public o1 evidence does not establish a specific MCTS + CriticGPT architecture.
| Question | Representative work | What is controlled? |
|---|---|---|
| Expose two modes | Qwen3 (May 2025) | Prompt / template + budget |
| Learn when to use each | Thinkless; AdaptThink (2025) | Mode-selection policy |
| Predict required budget | Sonata (ICLR 2026) | Compute allocation before reasoning |
| Make future direct answers better | System 2 distillation (2024) | Parameters, using selected outputs |
| Check simple-task behavior | S1-Bench (IJCAI 2026) | Evaluation, not a controller |
Research selection checked on 2026-11-15. Versions and primary sources are recorded in the reading page.
Qwen3 Technical Report · Thinkless: LLM Learns When to Think · AdaptThink: Reasoning Models Can Learn When to Think · Adaptive Thinking: Large Language Models Know When to Think in Latent Space · Distilling System 2 into System 1 · Exploring the System 1 Thinking Capability of Large Reasoning Models
The May 2025 report provides a controllable interface.
Mode availability alone does not demonstrate an optimal learned router.
§4.3: thinking-mode fusion. §4.7: budget analysis. These claims refer to the original unified-mode version.
| Table 1 · 1.5B | Baseline → Thinkless accuracy | Mean response tokens |
|---|---|---|
| GSM8K | 83.47% → 84.18% | 1919 → 624 |
| MATH-500 | 86.08% → 81.84% | 5675 → 2555 |
Token savings can accompany an accuracy loss. Thinking frequency, token count and wall time are different metrics.
Cold start: a long-thinking model may never sample direct answers without explicit exploration.
Importance sampling: correct for the sampling distribution while learning from both modes.
The performance constraint is an optimization target—not a proof that every future answer is correct.
DeepSeek-R1-Distill-Qwen-1.5B · original Thinking → AdaptThink · Table 1
| Dataset | Accuracy (%) | Response tokens | NoThinking share |
|---|---|---|---|
| GSM8K | 79.0 → 83.1 | 978 → 480 | 86.9% |
| MATH500 | 80.6 → 82.0 | 4887 → 1782 | 76.8% |
| AIME 2024 | 29.4 → 31.0 | 12073 → 6679 | 40.4% |
Three-dataset macro average: +2.4 pp accuracy; −53.0% response length.
16K evaluation context; temperature 0.6. AIME: 16 samples per problem, averaged—not best-of-16. Token reduction is not measured wall-clock speedup.
Self-consistency is a proxy for thinking need, not a truth certificate.
“Latent space” here supplies features to a budget predictor. It does not mean all reasoning becomes a Coconut-style latent chain.
20–60% fewer thinking tokens at the same accuracy across reported model / task settings.
Models include Qwen3-8B / 32B, GPT-OSS-120B, Qwen3-235B-A22B.
Tasks include AIME25, GSM8K, MATH500, GPQA and LiveCodeBench.
Not a universal per-query guarantee.
Not a tool-use success benchmark.
Not the same as measured latency.
Evaluate calibration cost, routing overhead, difficult-tail failures and actual task quality.
Version frozen to the conference proceedings; later author summaries report different experiment ranges.

4.6 Coordination and Collaboration 4.6.1 Workflows and Coordination























4.7 Learning and Evaluation 4.7.1 Feedback and System Optimization

Better search outputs may teach a better policy—if supervision is useful and transferable.
Possible failure: biased checks, copied errors, no useful reward, distribution shift.
Stop or revise: no held-out gain, worse reliability, excessive cost, exhausted evidence.
An iterative procedure is not evidence of unlimited improvement or inevitable superintelligence.








Yu et al. (2024): several methods transfer useful behavior; transfer is task-dependent.
Self-consistency filtering is not formal verification.
This experiment removes intermediate reasoning from the training target; other distillation designs can retain it.
Llama-2-70B-chat · GSM8K · k = 1 · original Table 4
| Condition | Accuracy | Generated tokens |
|---|---|---|
| Direct few-shot (no CoT) | 7.58% | 57 |
| Zero-shot CoT teacher | 52.77% | 270 |
| Answer-only CoT distillation | 7.13% | 18 |
Fewer output tokens did not preserve the teacher’s math accuracy.
This is one base model and one self-supervised answer-only recipe. It does not show that all reasoning distillation is impossible.
IJCAI 2026: evaluate simple tasks, response cost and consistency across repeated samples.
| Table 3 | pass@1 | acc@5 (all 5 correct) | Mean tokens |
|---|---|---|---|
| DS-R1-1.5B | 81.47% | 54.50% | 489.54 |
| DS-R1 (671B) | 100.00% | 100.00% | 646.40 |
Good hard-task scores do not settle simple-task reliability or efficiency.
T=0.6; top-p=0.95; k=5; max 10K. Paper-defined metrics use formatting checks and GPT-4o judging. “System 1” is a benchmark label, not a diagnosis of a mental mechanism.
| Compare on the same tasks | Report together |
|---|---|
| Always direct / always deliberate | Correctness, hard-constraint violations |
| Fixed budget / learned allocation | Tokens, actual latency, tail latency |
| Reason only / retrieve evidence | Tool cost, evidence gained, retries |
| Before / after training | Held-out transfer, training cost, regressions |
Count routing, failed attempts and tool calls. Preserve the same checking rules.
A policy that refuses every task is cheap; report coverage as well as accepted-case accuracy.
Tools, code, memory and retrieval provide operations for the controller.
Next: divide work across agents and check the joint result.
A successful local step is not enough: evaluate the complete task, including handoffs and failed attempts.

