Technical Summary: Language-Induced Priors for Domain Adaptation

Problem Statement

The paper addresses a fundamental paradox in Domain Adaptation (DA) within the cold-start regime. When a target agent (e.g., a new machine or patient) has scarce data ($N_0 \approx 0$), standard statistical methods struggle to distinguish relevant source domains from irrelevant ones. Indiscriminately pooling data from all available sources often leads to negative transfer, where irrelevant data biases the model and degrades performance below that of a model trained solely on the limited target data.

The core challenge is circular: to identify relevant sources, one needs a reliable estimate of the target parameters ($\theta_0$), but estimating $\theta_0$ requires knowing which sources to borrow from. The authors argue that while numerical target data is scarce, contextual information in the form of expert textual descriptions (e.g., "the machine is vibrating abnormally") is often available but overlooked. This paper proposes leveraging these descriptions to break the cold-start deadlock.

Methodology: Language-Induced Prior (LIP) and EM

The authors propose a probabilistic framework that integrates Language-Induced Priors (LIP) into an Expectation-Maximization (EM) algorithm for Multi-Source Domain Adaptation (MDA).

1. Problem Formulation

The setting involves $K+1$ domains: one target ($k=0$) and $K$ sources ($k=1, \dots, K$). Each domain $k$ is governed by latent parameters $\theta_k$. The goal is to estimate $\theta_0$.
The authors model the relationship between sources and the target using a Bayesian hierarchical model:

• A latent indicator $c_k \in \{0, 1\}$ determines if source $k$ is relevant ($c_k=1$) or irrelevant ($c_k=0$).
• If relevant, $\theta_k \sim \mathcal{N}(\theta_0, \tau^2 I)$.
• If irrelevant, $\theta_k \sim \phi_{\text{null}}$ (a noise component).

2. Language-Induced Prior (LIP) Construction

To resolve the circularity of the cold start, the framework introduces a prior $\pi_k = P(c_k=1)$ derived from a textual description of the target domain ($h$).

• Elicitation: Instead of asking an LLM for direct probabilities (which can be unreliable), the authors use subgroup preference elicitation. The LLM is presented with a subgroup of source datasets and a target description, then asked to select the most relevant source (or none).
• Modeling: These choices are modeled using a conditional logit model with an outside option (McFadden, 1974). The LLM's choice probability is:
  $$p(k_m | S_m) = \frac{e^{\alpha_{k_m}}}{e^{\alpha_0} + \sum_{j \in S_m} e^{\alpha_j}}$$
  where $\alpha_k$ are latent variables and $\alpha_0$ is the threshold for the null option.
• Optimization: The LIP parameters ($\pi_k = \sigma(\alpha_k)$) are estimated via Empirical Bayes by maximizing the likelihood of the LLM's responses, regularized to prevent overfitting.

3. LIP-aided EM Algorithm

The LIP is integrated into an EM algorithm to jointly estimate $\theta_0$ and the relevance indicators $c$.

• E-Step: Computes the posterior probability (weight) $w_k^{(t)} = P(c_k=1 | D_k, \theta^{(t)})$. This weight combines the LIP ($\pi_k$) with the empirical evidence (log-likelihood ratio between the relevant and null models).
  $$w_k^{(t)} = \sigma\left( \log \frac{p(D_k | c_k=1, \theta^{(t)})}{p(D_k | c_k=0)} + \log \frac{\pi_k}{1-\pi_k} \right)$$
  To handle computational intractability, the relevant likelihood is approximated using a second-order Taylor expansion of the log-likelihood.
• M-Step: Updates $\theta$ by maximizing a weighted log-likelihood. For low-dimensional problems, this admits a closed-form update (a precision-weighted blend of source MLEs). For high-dimensional models (e.g., neural networks), gradient ascent is used.
• Bayesian Tempering: To prevent early iterations from being dominated by noisy likelihood ratios when $\theta^{(t)}$ is poor, a tempering schedule $\beta_k^{(t)}$ is applied. This suppresses the data-driven likelihood ratio in early steps, allowing the LIP to guide the selection initially, with the weight shifting to data as samples accumulate.

Key Contributions

1. Language-Induced Prior (LIP): The first work to define and leverage LIPs for identifying source relevance in DA. It translates semantic descriptions into a choice model quantifying source relevance, guiding DA when target signals are weak.
2. LIP-aided EM: A Bayesian hierarchical framework solved via EM that allows expert intuition (via LIP) to guide source selection in cold-start scenarios while automatically refining these choices as data accumulates.
3. Theoretical Guarantees:
   • Oracle Performance: Under a correct LIP, the estimator's Mean Squared Error (MSE) matches an oracle cold-start MSE (which assumes knowledge of all relevant sources) up to exponentially small terms.
   • Asymptotic Consistency: Even if the LIP is misspecified, the algorithm remains asymptotically consistent as the target sample size ($N_0$) or source sample sizes ($N_k$) grow, serving as a safety net against incorrect priors.
4. Empirical Validation: The framework is validated on three distinct tasks:
   • Descriptive: Gaussian mean estimation (1-d and 2-d).
   • Predictive: C-MAPSS dataset (predicting engine core speed under degradation).
   • Prescriptive: MuJoCo Hopper (deep reinforcement learning with gravity variations).

Experimental Results

• Gaussian Estimation: LIP-aided EM successfully identified the relevant source and recovered the ground truth distribution, outperforming "Target Only" (high variance) and "Pooled" (negative transfer) baselines.
• C-MAPSS (Engine Degradation): In cold-start scenarios (e.g., predicting future speed with only 10-30% of the lifecycle data observed), LIP-aided EM significantly reduced RMSE compared to uniform priors and pooling. For instance, at 90% Remaining Useful Life (RUL), LIP-aided methods achieved RMSE ~14-15, while Target Only and Pooled methods failed with RMSE >30.
• MuJoCo Hopper: In a deep RL setting with limited target data ($N_0=128$), LIP-aided EM achieved substantially higher rewards than baselines. Notably, even a "False LIP" (generated by a model with reasoning errors) outperformed the uniform prior in the cold start, though it recovered slower than a correct LIP, demonstrating the algorithm's robustness.

Significance and Claims

The paper claims to bridge the gap between semantic reasoning and statistical learning. It posits that textual descriptions of a target environment carry information that statistical methods cannot recover from scarce data alone. By translating these descriptions into a Bayesian prior via LLMs, the method enables effective domain adaptation in regimes where traditional data-driven approaches fail.

The authors emphasize that while the quality of the LIP is crucial for optimal cold-start performance, the framework is theoretically robust: a correct LIP yields oracle-level precision, while any LIP (even incorrect ones) ensures asymptotic consistency. The work suggests a paradigm shift where contextual metadata (text) is treated as a primary modality for initializing statistical models, rather than an afterthought.

The paper does not claim to solve all domain adaptation problems but specifically targets the cold-start regime where target data is insufficient for reliable statistical inference, leveraging the often-overlooked resource of expert textual descriptions.