Thought Flow Nets: From Single Predictions to Trains of Model Thought

TL;DR

Thought Flow Nets introduce a dialectics-inspired, gradient-based self-correction mechanism that turns single-output predictions into iterative thought flows. By defining logits, a correctness score, and a gradient-driven update, the method enables sequential refinements in Transformer QA models, achieving up to $9.6\%$ absolute improvements in $F_1$ with an oracle stopping mechanism. Human evaluation indicates thought flows are perceived as more correct, understandable, and intelligent, while not increasing completion time. The work demonstrates a practical, task-agnostic correction framework with strong QA gains and rich qualitative patterns of self-correction, suggesting meaningful benefits for human-agent interaction and future learning-to-stop extensions.

Abstract

When humans solve complex problems, they typically create a sequence of ideas (involving an intuitive decision, reflection, error correction, etc.) in order to reach a conclusive decision. Contrary to this, today's models are mostly trained to map an input to one single and fixed output. In this paper, we investigate how we can give models the opportunity of a second, third and $k$-th thought. Taking inspiration from Hegel's dialectics, we propose the concept of a thought flow which creates a sequence of predictions. We present a self-correction mechanism that is trained to estimate the model's correctness and performs iterative prediction updates based on the correctness prediction's gradient. We introduce our method at the example of question answering and conduct extensive experiments that demonstrate (i) our method's ability to correct its own predictions and (ii) its potential to notably improve model performances. In addition, we conduct a qualitative analysis of thought flow correction patterns and explore how thought flow predictions affect human users within a crowdsourcing study. We find that (iii) thought flows enable improved user performance and are perceived as more natural, correct, and intelligent as single and/or top-3 predictions.

Figures (7)

• Figure 1: In contrast to the standard approach of mapping an input to an output in a single step (grey box), we propose a method that allows models to sequentially "reconsider" and update their predictions, i.e., the thought flow. In this (real) question answering example, the orange box marks our thought flow extension, which corrects a flawed answer in two steps.
• Figure 2: The steps of the prediction update scheme and their relation to the three moment of Hegel's Dialectics. The example shows the first answer change from \ref{['fig:teaser']}.
• Figure 3: Thought flows with different gradient scaling targets $\delta$ averaged over three seeds of a question answering model. Higher values for $\delta$ correspond to more aggressive decision changes. Without a stopping oracle that stops when the thought flow does no longer improve an answer (left), only $\delta=0.1$ provides consistently stable, but very small F1 improvements. With an oracle (middle), higher values for $\delta$ reach higher and faster F1 improvements up to $>$9%. Nearly all performance gains are achieved by the first decision change (right). y axes use a symlog scale. Improvements are reported as absolute F1 scores (not relative to the the base performance).
• Figure 4: User study interface showing the TF condition (ours).
• Figure 5: User study interface showing the top-3 condition.