Chain-of-Personalized-Reasoning (CoPeR)

• Chain-of-Personalized-Reasoning (CoPeR) is a framework that decomposes AI reasoning into sequential stages, explicitly incorporating user preferences.
• It employs modular architectures that integrate basic control, intent modeling, and latent preference inference for dynamic personalization.
• CoPeR enhances task performance through improved action accuracy and balanced optimization of correctness and personalized outputs.

Chain-of-Personalized-Reasoning (CoPeR) denotes a paradigm in artificial intelligence and machine learning by which agents, particularly large-scale embodied and LLMs, integrate user-specific preferences into sequential reasoning processes. CoPeR generalizes approaches where an agent’s action and inference pipeline is systematically decomposed into multiple stages—such as basic control, explicit intent modeling, and latent preference inference—with each stage explicitly grounded in observed user cues and updated preference models. This class of methodologies has emerged to address the limitations of conventional task- or correctness-oriented models, which often fail to adapt output to individual user requirements or preferences in dynamic, personalized, or “cold start” contexts (Zhang et al., 10 Dec 2024, Li et al., 30 Sep 2025).

1. Formalization and Theoretical Framework

CoPeR frameworks are distinguished by the explicit factorization of reasoning into a staged chain, with each stage conditionally dependent on both the state of the environment/task and the evolving user preference profile. This is codified in several representative systems as follows:

Let an agent’s action space be $$\mathcal{A}_{\text{existing}} = \{\text{Action}_i\mid \text{Task goal}\}$$. The CoPeR extension is

$$\mathcal{A}^{\text{coper}} = \{\text{Action}_i\mid \text{Task goal, User preference}\}.$$

The staged reasoning unfolds via functions

$$\begin{aligned} \text{Stage 1:}~& a_t = f_1(o_t,\,i,\,h_{t-1}) \ \text{Stage 2:}~& u = f_2(\{o_1\dots o_k, a_1\dots a_k\},\,i) \ \text{Stage 3:}~& r = f_3(\{v_j\},\,u) \end{aligned}$$

where $o_t$ is observation, $i$ is an instruction, $h_{t-1}$ is history, $u$ is an explicit (textual or symbolic) preference or requirement, and $\{v_j\}$ are candidate objects. The functional form enables modular inference, where each stage’s output conditions the next.

In the context of LLMs, CoPeR generalizes to a preference elicitation and alignment process. Given a global attribute set $\Theta=\{\theta_1,\ldots,\theta_d\}$ and instance-specific relevant set $\mathcal{F}(i)\subset\Theta$, a user’s hidden profile is $$\mathcal{P}_{p,i} = \{(\theta_j, v_j, w_j): \theta_j\in\mathcal{F}(i)\}$$, with value $v_j$ and importance $w_j$ on attribute $\theta_j$. The objective is to generate a response $r$ satisfying both

$$\text{Correct}(r,i)=1, \qquad \text{PrefAlign}(r,\mathcal{P}_{p,i}) = \sum_{\theta_j\in\mathcal{F}(i)}w_j\,g_j(r,v_j)$$

where $g_j$ quantifies alignment to each user preference (Li et al., 30 Sep 2025).

2. Modular Architectures and Workflow

CoPeR implementations instantiate a structured workflow, typically decomposed along the following axes:

Stage	Embodied Agents (Zhang et al., 10 Dec 2024)	LLM Personalization (Li et al., 30 Sep 2025)
Stage 1	Basic GUI/robotic navigation	Identify relevant preference attributes
Stage 2	Explicit user requirement inference	Just-in-time elicitation via questioning
Stage 3	Implicit personalized recommendation	Reasoning chain adaptation and personalized response

Embodied frameworks such as SmartAgent (Zhang et al., 10 Dec 2024) use a Perceiver module (vision-language backbone) for action and recommendation, and a Reasoner module for synthesizing explicit preferences. The workflow typically consists of initial navigation to a candidate set (item pool), inference of explicit user requirements from interaction history, and utility-maximizing recommendation or action selection conditioned on those preferences.

In LLMs, after preference-uncertainty assessment, the agent decides sequentially whether to ask preference-disambiguating queries or produce an output. The agent adaptively determines which attributes to elicit (via expected information gain), integrates received values, and then conditions the response-generation chain of thought on the updated user profile.

3. Datasets and Evaluation Methodologies

CoPeR research requires datasets that provide both granular action traces and explicit preference signals. In the embodied domain, SmartSpot (Zhang et al., 10 Dec 2024) is the first benchmark focused on embodied personalized tasks, constructed around a multi-channel real-world app. It provides:

• 144 episodes spanning 1,400 steps in 7 scenarios (single- and multi-channel)
• Per-step GUI screenshots, ground-truth actions, full interaction history, task instructions, and paired underlying requirements
• Partitioned protocol: GUI navigation, item pool selection, and personalized recommendation stages

Metrics include element-level action accuracy (Ele.Acc), step success rate (SSR), explicit preference alignment (Exp.Acc via embedding similarity), and implicit recommendation accuracy (Imp.Acc).

For LLM-based CoPeR, PrefDisco (Li et al., 30 Sep 2025) defines a protocol for transforming static benchmarks into interactive personalization tasks by sampling psychologically-grounded user personas and sparse preference profiles, and using LLM-based rubrics for evaluating preference alignment. Evaluation measures include overall accuracy, normalized preference alignment, number of preference-eliciting queries, and specific task-level breakdowns.

4. Training Objectives and Optimization

CoPeR implementations introduce multi-stage or multi-objective loss functions to reflect their sequential architecture:

• Embodied Action Loss: $$L_1 = -\sum_{t\in \text{GUI}} \log p(a_t^\star | o_t, i, h_{t-1})$$
• Explicit Preference Loss: $L_2 = 1 - \cos(\phi(\hat u), \phi(u^\star))$ for semantic similarity between predicted and ground-truth explicit requirements
• Personalized Recommendation Loss: $$L_3 = -\sum_j \left[y_j \log p(r_j = \text{Yes} | v_j, \hat u) + (1-y_j) \log p(r_j = \text{No} | v_j, \hat u)\right]$$

These are minimized in GPT-style auto-regressive training. LoRA tuning (low-rank adaptation) is applied to both visual and language layers using AdamW optimization. Joint versus stage-wise optimization strategies have differential effects on explicit versus implicit preference alignment, revealing trade-offs (Zhang et al., 10 Dec 2024).

For LLMs evaluated under PrefDisco, the principal optimization goal is to improve preference alignment ($\text{PrefAlign}$) subject to correctness constraints, utilizing decision policies for preference elicitation and adaptation of internal reasoning chains to maximize alignment scores.

5. Empirical Results and Failure Modes

Empirical results on SmartSpot (Zhang et al., 10 Dec 2024) show that SmartAgent, which implements a three-stage CoPeR chain, achieves highest element accuracy (0.64) and step success rate (0.50) among tested agents, with explicit preference accuracy (0.71) and implicit recommendation accuracy (0.24). Multi-channel scenes exhibit the largest margins for preference-related metrics, and zero-shot generalization is notably lower for embodied control than for preference inference. End-to-end training slightly boosts embodied control and implicit recommendation, but at a modest cost to explicit alignment.

For LLM systems, evaluations in PrefDisco (Li et al., 30 Sep 2025) demonstrate that 29.0% of naive personalization attempts decrease alignment versus generic responses, and average question count is low (1.48 / 5-turn budget), limiting potential for improved alignment. Accuracy of personalized responses can decrease, especially on formal domains (e.g., math tutoring tasks with –12.1% drop), due to over-correction and brittleness when forced to adapt reasoning chains. These findings establish that naïve or uncalibrated CoPeR pipelines can degrade both correctness and utility for end-users.

6. Open Research Challenges and Generalization

Several key challenges influence the generalizability and maturation of CoPeR:

• Ambiguity in user instructions and difficulty of reliable measures for “user satisfaction”
• Dynamic, online updating of user preference models as interactions unfold
• Multi-user/multi-profile adaptation, particularly for shared devices or group settings
• Evaluation beyond binary accuracy, incorporating satisfaction signals (e.g., dwell time, engagement surveys)
• Robustness and safety with respect to over-personalization or unintended consequences of preference adaptation
• Integration of multi-dimensional, attribute-level reward models and reinforcement learning that jointly optimizes for correctness and preference satisfaction (Li et al., 30 Sep 2025)

Extending CoPeR effectively into new domains requires defining modular multi-stage task decompositions, assembling or adapting datasets to supply both action traces and user preference ground-truths, and designing architectures whose modules can be jointly or sequentially optimized for personalization-aware losses (Zhang et al., 10 Dec 2024).

7. Relationship to Chain-of-Thought and Personalization Paradigms

CoPeR generalizes the concept of chain-of-thought (CoT) reasoning by requiring not only that models expose internal reasoning but also that they select, adapt, and possibly re-weight reasoning pathways based on inferred user-specific goals or constraints. In contrast to standard CoT, which is mostly static and correctness-oriented, CoPeR introduces a deliberate, user-centric adaptation loop—eliciting information where uncertainty is highest, explicitly modeling multi-attribute preferences, and conditioning output on a dynamic profile. This establishes CoPeR as a distinct research frontier that unifies preference elicitation, sequential inference, and adaptive reasoning for personalized, interactive systems (Li et al., 30 Sep 2025, Zhang et al., 10 Dec 2024).