Recon : Reconstruction-Guided Reasoning Synthesis for User Modeling

Alan Zhu Mihran Miroyan Carolyn Wang Andrew Zhou Lisa Dunlap Narges Norouzi Joseph E. Gonzalez University of California, Berkeley · 2026

Read Recon : Reconstruction-Guided Reasoning Synthesis for User Modeling by Alan Zhu Mihran Miroyan Carolyn Wang Andrew Zhou Lisa Dunlap Narges Norouzi Joseph E. Gonzalez University of California, Berkeley (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-26969.txt.

Introduction

Zhu and six colleagues build Recon, a method that scores reasoning traces for user modeling by their power to predict actions. Older methods condition traces on both context and action, guaranteeing justification but not true causal reasoning. Recon instead asks a reconstruction model to predict the action from context plus candidate reasoning, and fidelity sets trace quality. Recon wins 54.7 percent of comparisons against Backward Synthesis and up to 70 percent downstream, and its traces transfer across models.

Catalog record

CREATORAlan Zhu Mihran Miroyan Carolyn Wang Andrew Zhou Lisa Dunlap Narges Norouzi Joseph E. Gonzalez University of California, Berkeley
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63580
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.26969

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Recon : Reconstruction-Guided Reasoning Synthesis for User Modeling Alan Zhu Mihran Miroyan Carolyn Wang Andrew Zhou Lisa Dunlap Narges Norouzi Joseph E. Gonzalez University of California, Berkeley Abstract User modeling aims to use language models (LMs) to mimic an individual’s behavior from a corpus of past context–action pairs (e.g., conversation turns), enabling the simulation of users in settings like behavioral science, human–AI collaboration, and market research. Recent approaches augment these corpora with synthesized reasoning traces, typically generated by conditioning on both context and action. However, such conditioning constitutes post-hoc rationalization rather than reasoning: the trace is guaranteed to justify the action, but may not encode the underlying latent causal decision paths. We propose Recon , which uses action reconstruction to score reasoning traces by their predictive power: given a context and candidate reasoning, a reconstruction model predicts the action, and reconstruction fidelity determines reasoning quality. Across four domains, Recon achieves a 54.7% win rate over Backward Synthesis, a standard post-hoc rationalization baseline. Further, we find that training a reasoning synthesis model with rewards derived from Recon improves downstream user modeling performance, achieving a win rate of up to 70.0% over baselines. We further show that Recon -synthesized reasoning transfers across models, and improves user modeling beyond the reconstruction model. Our work demonstrates that post-hoc rationalization is insufficient for reasoning synthesis, and that useful and interpretable reasoning should naturally elicit the action from the context. 1 Introduction Figure 1: Recon pipeline. Given a context–action pair (c,a^{*}) (1), we sample N=4 candidate rationalizations from the reasoning model M_{r} (2). Each candidate is evaluated by using the action model M_{a} to reconstruct the action from the context c and candidate reasoning \hat{r}_{i} , and measuring its alignment with the ground-truth action a^{*} (3). The resulting scores are used either to select rationalizations for Recon -Select synthesis or as reward signals for RL training in Recon -GRPO. A language model (LM) that understands how a user thinks is better able to predict how a user acts . LMs are increasingly used to simulate individual users in applications such as customer feedback [ 13 ] , human subjects research [ 9 ] , and human-AI collaboration [ 27 ] . In these settings, the goal is to predict a user’s action in a given context from past interactions, which may include preference labels [ 21 , 25 ] , interview responses [ 19 ] , or unstructured natural language interactions [ 3 , 13 , 15 ] . For instance, the context may consist of prior conversation turns, and the action may be the user’s next-turn response. Recent work synthesizes reasoning traces that explain the reasoning behind an action and provide information beyond the original context–action pair, improving downstream action prediction generalization [ 11 , 13 , 15 ] . These reasoning synthesis methods draw on approaches used in verifiable domains such as math, multi-hop QA, and coding [ 6 , 22 , 30 ] , where models are prompted to generate reasoning linking the context to the action. However, these synthesized traces are post-hoc rationalizations prone to logical leaps and arbitrary decisions. In logic-heavy tasks, such rationalizations still produce a logically sound chain from context to answer. However, in open-ended free-response settings like conversations, the rationalization does not necessarily lead naturally to the observed action. As a simple example, suppose the context involves someone at a cafe, and the individual’s action is ordering a glass of orange juice. A possible rationalization is, “I was feeling thirsty.” Although consistent with the action, this rationale does not explain the specific choice of orange juice, since thirst would more naturally support ordering water. A better rationale is, “I like orange juice,” which more directly explains the user’s observed action. Indeed, prior work in cognitive science shows that human rationalizations often diverge from underlying decision processes [ 16 ] . An alternative to post-hoc rationalization generation is to condition only on the context and retain candidate reasoning traces that produce the correct outcome, i.e., rejection sampling. This approach can be effective in settings with small output spaces where it can be expected that with enough samples, the observed action will be covered, such as binary preferences [ 21 ] . However, in open-ended natural language settings, where actions are high-dimensional, simple rejection sampling becomes infeasible. Further, prior work has explored reasoning training methods [ 26 ] , where a model generates both reasoning and action and is rewarded based on alignment of the generated action with the ground-truth. However, the reward depends only on the action, so the resulting reasoning traces need not be interpretable or useful for understanding user behavior, as shown in Section 5 . This leaves a gap: how can we synthesize reasoning traces that are both aligned with observed actions and useful for downstream prediction? To address this, we propose Recon , which evaluates candidate rationalizations based on action reconstruction . We first synthesize a candidate reasoning trace conditioned on both context and ground truth action. Then, given the same context and the synthesized reasoning trace, we prompt an LM to generate a reconstructed action and measure its alignment with the ground truth. By generating and ranking multiple candidates, Recon selects rationalizations that predict user behavior, generating useful, interpretable reasoning traces for in-context learning methods ( Recon -Select). Moreover, the rankings induced by Recon over a set of candidates can be used to obtain a reward signal for training a reasoning synthesizer ( Recon -GRPO). The end-to-end Recon pipeline is shown in Figure 1 . We evaluate Recon across four domains: US Supreme Court oral arguments, UK Prime Minister’s Questions, podcasts, and Reddit conversations. For all tasks, context consists of past turns in the conversation, and action is a spoken or written response. We compare against the reasoning synthesis method used in 13 and 30 . Methods are evaluated by synthesizing reasoning traces to augment a retrieval corpus used by a fixed retrieval-augmented generation (RAG) pipeline to generate actions for test contexts. An LM-judge performs pairwise comparison of actions generated from corpora augmented by different synthesis methods for alignment with the ground truth action; we then report win rates after excluding ties. Across four domains, Recon -Select improves over existing baselines with overall win rates of 54.7% and 53.5% using Qwen3-8B and GPT-5-mini, respectively, while Recon -GRPO further improves win rates up to 70%. Our contributions are as follows: 1. We introduce Recon , a method that uses action reconstruction as a criterion for evaluating synthesized reasoning in user modeling settings. 2. We use Recon to synthesize reasoning via both training-free and training-based methods, achieving win rates over existing methods of up to 70% with training and consistently above 50% with prompt-based synthesis. 3. We demonstrate that reasoning synthesized using Recon is transferable and interpretable, improving models not used in the synthesis pipeline. 2 Related Work 2.1 User Modeling and Personalization User modeling, the task of modeling an individual user’s behavior from past interactions, is closely related to personalization, which broadly aims to align models with user preferences. Much of LM-based personalization assumes access to pairwise preference labels (c,o^{(1)},o^{(2)},y) , where y indicates which output o^{(i)} a user prefers in context c . Personalized-RLHF

Read the complete text