ChatLLM Network: More brains, More intelligence

• Single dialogue models may vary answers under the same prompt and provide one-sided responses, limiting comprehensiveness.Source: §1 · p. 2
• ChatLLM combines forward aggregation with language-based feedback to update model reasoning, while dropout manages inputs to prevent information overload.Source: §1 · p. 2
• The network lets multiple dialogue models interact, provide feedback, and think together, and can apply to different types of dialogue models.Source: §1 · p. 2
• Experiments on digital mode classification and sentiment reversal show significant problem-solving enhancements over the vanilla model.Source: §1 · p. 3

2.1 Large Language Models

• Transformer pretraining enabled large-scale unsupervised text training; BERT and later models advanced NLP, while GPT-3 [Brown et al., 2020] showed zero-shot ability.Source: §2.1 · p. 3

2.2 ChatGPT

• ChatGPT is described as instruction-tuned GPT-3 trained with human feedback, producing more aligned and helpful responses to user instructions.Source: §2.2 · p. 3
• HuggingGPT (Yongliang et al. [2023]), AutoGPT (Toran [2023]), generative agents (Joon et al. [2023]), and CAMEL (Guohao et al. [2023]) explore cooperation; unlike CAMEL’s task decomposition, this work encourages shared reflection.Source: §2.2 · p. 3

2.3 Improving Language Models via Feedback

• SELF-REFINE [Madaan et al., 2023] uses iterative self-feedback and refinement, while self-ask [Press et al., 2022] targets compositional reasoning.Source: §2.3 · p. 3

3.1 Network Architecture

• The network has n − 1 fully connected layers followed by one aggregation layer whose leader combines prior inputs and produces the final output.Source: §3.1 · p. 4

3.2 Forward-aggregation Mechanism

• Each model receives the question and outputs from earlier layers, so later models can identify useful highlights and aggregate viewpoints that may offset individual deviations.Source: §3.2 · pp. 4–5

3.3 Language Based Backpropagation Mechanism

• After comparing the final output with the answer, feedback prompts preserve correct reasoning or ask incorrect models to diagnose and revise it across network layers.The stated aim is to correct incorrect suggestions while keeping models with correct reasoning stable.Source: §3.3 · p. 6

3.4 Dropout Mechanism

• To limit input volume, each model randomly receives messages from only a limited number of other models, with ρ denoting the rate of selected models.Source: §3.4 · p. 7

3.5 Network Optimization

• Training examples are processed sequentially, with language-based updates applied per sample; training stops at a preset iteration count or when performance no longer improves.Source: §3.5 · p. 7

4.2 Sentiment Reversal Experiment

• Sentiment reversal tests whether the network can improve a traditional NLP task, using feedback to intensify reversed sentences and a separate ChatGPT judge to compare emotional intensity.The study generates 60 emotionally biased sentences; the baseline uses self-feedback, whereas the network uses backpropagation for feedback.Source: §4 · p. 7, §4.2 · p. 9

4 Experiments

• The network uses three ChatGPTs in its first layer and one ChatGPT in its second layer, with the latter consolidating employee outputs.Experiments use ChatGPT3.5 as the basic member.Source: §4 · p. 7, §4.2 · p. 9

4.1 Digital Mode Classification

• Digital mode classification tests inductive learning from scratch by asking ChatGPT to classify three-dimensional vectors according to the position of their largest value.The task rules are not part of ChatGPT's pre-existing knowledge; evaluation uses 30 manually designed challenging samples, without feedback during testing.Source: §4.1 · p. 8
• ChatGPT network members achieve significantly higher accuracy than the baselines, and the leader's output is more accurate than other members.The leader's accuracy variance is generally similar to ChatGPT-ensemble and lower than other baselines, which the authors interpret as relatively stable output.Source: §4.1 · p. 8, Table 1 · p. 8
• Accuracy curves mostly peak early; ChatGPT curves often become unstable afterward, while the network's criteria grow more complex in a pattern resembling overfitting.The authors say early stopping is necessary to achieve the best results.Source: §4.1 · p. 8, Figure 6 · p. 9

4.2 Sentiment Reversal Experiment

• The ChatGPT network significantly outperforms standalone ChatGPT on sentiment reversal with feedback, while its no-feedback advantage is marginal.The authors attribute the no-feedback advantage to summarizing information.Source: Table 2 · p. 9, §4.2 · p. 9
• Examples show network-generated reversals using more intense emotional adjectives and richer vocabulary than ChatGPT outputs.The examples are presented in Table 3.Source: §4.2 · p. 10, Table 3 · p. 10

• Limited dialogue-model capability, especially with large numerical inputs, means the network may not be superior in every scenario; lacking an effective inter-model communication mechanism also restricts experiments to few network members.Source: §5 · p. 10

• The ChatLLM network lets multiple dialogue-based language models interact and provide feedback, consolidating their distinct perspectives through a separate model for decision-making.Source: §6 · p. 10
• The network's optimization uses a language-based backpropagation mechanism.Source: §6 · p. 10
• Experiments on two tasks demonstrate the network's effectiveness and superiority.Source: §6 · p. 10
• The authors plan a global strategy assigning distinct identities and exclusive tasks to models to enhance inter-model communication traceability.Source: §6 · pp. 10–11

A.1 Forward-aggregation

• The three ChatGPT instances offer differing classifications of the same data point, with some favoring the first category and another favoring the second.Source: §A.1 · p. 12
• A separate leader presents the instances’ responses together and weighs their rationales before making its own classification.Source: §A.1 · p. 12
• The leader leans toward the second category because it treats the second feature value as more important than the third.Source: §A.1 · p. 12

A.2 Backpropagation

• After being told its guess was wrong, ChatGPT1 is asked to speculate about the reason and update its classification criteria using another person’s reasoning.Source: §A.2 · p. 13
• The leader’s reasoning identifies disagreement among the members and favors the second category based on the second feature value.Source: §A.2 · p. 13
• ChatGPT1 revises its answer to the second category, citing the second feature value as the highest and an important classification factor.Source: §A.2 · p. 13