💬 NLP

Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling

The paper introduces Recon, a method that improves user modeling by scoring and synthesizing reasoning traces based on their ability to predictively reconstruct user actions, thereby moving beyond ineffective post-hoc rationalization to capture genuine causal decision paths.

Original authors: Alan Zhu, Mihran Miroyan, Carolyn Wang, Andrew Zhou, Lisa Dunlap, Narges Norouzi, Joseph E. Gonzalez

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alan Zhu, Mihran Miroyan, Carolyn Wang, Andrew Zhou, Lisa Dunlap, Narges Norouzi, Joseph E. Gonzalez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to be a specific person—let's say, a grumpy but brilliant Supreme Court Justice or a chatty podcast host. You have a library of their past conversations (the "context") and what they actually said next (the "action"). Your goal is to make the robot predict what they would say in a new situation.

To help the robot understand why the person said what they said, researchers usually try to generate a "thought process" or a "reasoning trace" to go along with the conversation.

The Problem: The "Hindsight" Trap
Most current methods work like a lawyer writing a closing argument after the verdict is already known. They look at the context and the final answer, then write a story that makes the answer look logical.

The Flaw: This is called "post-hoc rationalization." It's like saying, "I ordered orange juice because I was thirsty." While true, it doesn't explain why they chose orange juice specifically instead of water. A better reason might be, "I love the taste of orange juice."
The Result: The robot learns to write stories that justify the answer, but these stories don't actually capture the real, hidden thought process that led to the decision. It's a "fake" reason that just happens to fit.

The Solution: RECON (Reconstruction-Guided Reasoning)
The authors of this paper propose a new method called RECON. Instead of just asking the robot to write a story that fits the answer, they use a "test drive" approach.

Think of it like a cooking challenge:

The Setup: You have a recipe (the context) and a finished dish (the user's action).
The Guess: You ask a chef (the reasoning model) to write down their thought process for making that dish.
The Test (The "Reconstruction"): You take that written thought process and give it to a different chef (the action model) who hasn't seen the original dish. You ask them: "Based only on these thoughts and the ingredients, what dish would you make?"
The Score: If the second chef makes a dish that tastes exactly like the original, the thought process was good. If they make something totally different, the thought process was a bad guess, even if it sounded logical on paper.

How They Used It
The researchers used this "test drive" in two ways:

RECON-Select (The Filter): They generated four different "thought processes" for a single conversation. They ran the test drive on all four. The one that allowed the second chef to recreate the original dish most accurately was selected as the "best" reasoning. They used this to build a better training library.
RECON-GRPO (The Coach): They used the score from the test drive as a "reward" to train the robot chef directly. If the robot wrote a thought process that led to a perfect reconstruction, it got a high score. If not, it learned to try again.

The Results
They tested this on four very different types of conversations:

U.S. Supreme Court oral arguments (formal, legal).
UK Prime Minister's Questions (political debates).
Podcasts (casual interviews).
Reddit threads (internet arguments).

What They Found:

Better than the old way: The RECON method beat the standard "hindsight rationalization" method about 55% to 70% of the time.
Real reasoning vs. Fake justification: Simply training a model to get the right answer didn't work well (it only won 38% of the time). The model learned to "cheat" by memorizing the answer rather than understanding the user's mind. But when they used the RECON "test drive" to select or train the reasoning, the model actually learned how the user thinks.
It works on other robots: The "thought processes" created by RECON worked well even when used by a different AI model than the one that created them. This proves the reasoning was capturing the user's personality, not just the quirks of the AI that wrote it.

In Short
The paper argues that to truly simulate a human, you can't just ask an AI to write a story that explains an answer after the fact. You have to check if that story is strong enough to produce the answer on its own. RECON is the tool that checks that strength, leading to much more accurate and realistic user simulations.

Technical Summary: RECON: Reconstruction-Guided Reasoning Synthesis for User Modeling

Problem Statement

User modeling aims to simulate individual user behavior using language models (LMs) by predicting actions (e.g., conversation turns) based on past context–action pairs. Recent approaches attempt to improve generalization by augmenting these corpora with synthesized reasoning traces. However, existing methods typically rely on post-hoc rationalization, where a model generates a reasoning trace conditioned on both the context and the ground-truth action.

The authors argue that this approach is fundamentally flawed for user modeling. While such traces are guaranteed to justify the observed action, they often constitute logical leaps or arbitrary decisions that do not encode the underlying latent causal decision paths. In open-ended natural language settings, a rationalization might be consistent with an action (e.g., "I was thirsty" justifying ordering orange juice) without explaining the specific choice (e.g., "I like orange juice" vs. water). Consequently, these traces fail to capture the user's true latent reasoning ( $r^*$ ), limiting their utility for downstream prediction. Furthermore, simple rejection sampling (generating reasoning until the correct action is produced) is infeasible in high-dimensional, free-form natural language spaces.

Methodology: RECON

To address the gap between post-hoc rationalization and true reasoning, the authors propose RECON (Reconstruction-Guided Reasoning Synthesis). The core insight is that useful reasoning should naturally elicit the action from the context, rather than merely justifying a pre-determined outcome. RECON evaluates candidate reasoning traces based on their action reconstruction fidelity.

The RECON pipeline operates as follows:

Candidate Generation: Given a context $c$ and ground-truth action $a^*$ , a reasoning model $M_r$ generates $N$ candidate reasoning traces $\{\hat{r}_i\}$ . These are initially conditioned on $(c, a^*)$ , similar to standard rationalization.
Action Reconstruction: An action model $M_a$ is prompted with the context $c$ and a candidate reasoning trace $\hat{r}_i$ (without the ground-truth action) to generate a reconstructed action $\hat{a}_i = M_a(c, \hat{r}_i)$ .
Scoring: The quality of $\hat{r}_i$ is determined by the alignment between the reconstructed action $\hat{a}_i$ and the ground-truth action $a^*$ . This is evaluated by an LM judge across three dimensions: style, intent, and values.
Application: The reconstruction scores are used in two distinct ways:
- RECON-Select (Training-Free): The method samples multiple candidates and selects the one with the highest reconstruction alignment score to augment the retrieval corpus for in-context learning.
- RECON-GRPO (Training-Based): The reconstruction alignment score serves as a reward signal for training a reasoning synthesizer using Group Relative Policy Optimization (GRPO). Here, the reasoning model $M_r$ acts as an "advisor" to a frozen action model $M_a$ , learning to produce reasoning traces that facilitate accurate action reconstruction.

Key Contributions

RECON Framework: The introduction of a method that uses action reconstruction as a tractable surrogate for rejection sampling to evaluate synthesized reasoning in user modeling.
Dual Synthesis Strategies: Demonstration of RECON's effectiveness through both training-free selection (RECON-Select) and training-based optimization (RECON-GRPO).
Transferability and Interpretability: Evidence that reasoning traces synthesized via RECON are transferable across different model families and improve user modeling performance beyond the specific reconstruction model used during synthesis.

Experimental Results

The authors evaluated RECON across four domains: U.S. Supreme Court oral arguments, U.K. Prime Minister's Questions (PMQ), podcasts, and Reddit conversations. The evaluation involved a retrieval-augmented generation (RAG) pipeline where an action model predicts user responses using augmented context-action pairs.

Performance vs. Baselines: RECON-Select consistently outperformed the standard Backward Synthesis baseline (post-hoc rationalization).
- Using Qwen3-8B, RECON-Select achieved an overall win rate of 54.7%.
- Using GPT-5-mini, RECON-Select achieved an overall win rate of 53.5%.
- In contrast, the E2E-GRPO baseline (training a model to generate reasoning and action directly for action accuracy) performed poorly, achieving only a 38.4% win rate, suggesting that optimizing solely for action accuracy does not yield transferable reasoning.
Training with RECON Rewards: When RECON-GRPO was used to train reasoning models, win rates improved significantly, reaching up to 70.0% (specifically for smaller models like Qwen3-4B on the PMQ domain).
Cross-Model Transfer: Experiments swapping reasoning and action models (e.g., using GPT-5-mini reasoning with Qwen3-8B action generation) showed that RECON-Select reasoning traces remain useful, achieving a 52.4% win rate. This indicates the method recovers latent user characteristics rather than model-specific artifacts.
Dimensional Analysis: Improvements were consistent across style, intent, and values, indicating that RECON captures high-level user characteristics, not just surface-level linguistic patterns.

Significance and Claims

The paper claims that post-hoc rationalization is insufficient for reasoning synthesis in user modeling. The authors argue that useful and interpretable reasoning must be evaluated by its causal consequences—specifically, its ability to naturally elicit the observed action from the context—rather than its surface consistency with the action.

RECON provides a "tractable instantiation" of this principle. By prioritizing reasoning that successfully reconstructs the action, the method generates traces that better approximate the user's latent reasoning process ( $r^*$ ). The results suggest that this approach improves the fidelity of user simulation in noisy, high-dimensional natural language settings, offering a path toward more interpretable and predictive user models without relying on controlled preference labels. The authors note limitations, including the computational cost of the reconstruction step and the fact that RECON still operates on rationalizations conditioned on the ground-truth action, leaving the generation of reasoning conditioned only on context as a direction for future work.