Certified Behavioral Retention for Continual Adaptation with Held-Out Anchors

Ahanaf Hasan Ariq, Abrar ShahidarXiv 2026 · arXiv

TL;DR

The paper introduces Certified Anchor-Based Adaptation (CABA), a post-update "gate" that ensures model updates don't deviate too far from previously learned behaviors. By testing updates against a held-out set of "anchor" observations and using parameter interpolation, CABA provides a mathematically grounded way to trade off new learning (plasticity) against the retention of old skills (stability).

What Problem Does This Paper Address?

In the world of continual learning, agents (like LLMs or robots) face the stability-plasticity dilemma: how do you learn new tasks without "catastrophic forgetting" of old ones?

Most existing solutions, such as experience replay or parameter regularization, intervene during the training process. However, for a deployed system, an operator needs more than just a training heuristic; they need auditable evidence that a proposed update won't break existing, relied-upon behaviors before it is committed. The authors frame this as a "change-management" problem for foundation models, where updates must be certified against a pre-declared budget for behavioral drift.

Methods

CABA acts as a candidate-agnostic wrapper that sits between the update generator and the final deployment. Its workflow is as follows:

  1. Data Partitioning: Incoming data is split into a training set and a held-out anchor set. The anchors are never seen during the candidate's optimization.
  2. Interpolation Path: Given a current model ($\theta_{t-1}$) and a proposed candidate ($\hat{\theta}_t$), CABA defines a path of models $\theta(\alpha)$ by interpolating between them, where $\alpha \in [0, 1]$.
  3. Worst-Anchor Check: The gate evaluates the forward KL divergence between the old policy and the interpolated policy across all anchors. Crucially, it looks at the maximum (worst-case) divergence rather than the average.
  4. Selection: It selects the largest interpolation step $\alpha$ that stays within a pre-defined behavioral budget $\epsilon$.
  5. Certification: The authors provide a theoretical framework showing that if the anchor set sufficiently covers the state space (a $\rho$-net) and the policy is sufficiently smooth (Lipschitz continuous), the finite check on anchors extends to a uniform guarantee across the entire supported behavior.

Unlike many methods, CABA is label-free; it only needs to compare the output distributions of the two models on the same inputs, meaning it doesn't require ground-truth labels for the retention check.

Table 1: Relationship to representative retention mechanisms. “Post-candidate” means retention is checked after candidate generation.

Datasets

The authors evaluate CABA using a controlled study on the UCI Human Activity Recognition (UCI-HAR) dataset.

  • Setup: They treat the dataset as a non-stationary sensor-policy proxy, where the model must adapt to new subjects over time.
  • Architecture: They use a 561-dimensional feature vector (treated as a frozen representation) and a linear softmax head as the policy.

Key Insights and Ideas

  • Worst-Anchor vs. Average: Most methods focus on average performance. CABA’s focus on the worst-anchor makes it a "hard constraint" system. This is vital for safety; a model that is 99% fine but fails catastrophically in one specific, critical state is often unacceptable.
  • Auditable Receipts: CABA generates a machine-readable "receipt" (Table 3 in the paper) that logs the budget, the anchors used, the tested path, and the final decision. This makes model updates transparent and reversible.
  • Separation of Concerns: CABA doesn't replace training methods like Replay or GEM. Instead, it provides an independent layer of verification. You can use any "candidate generator" you want; CABA simply decides how much of that candidate is safe to commit.
  • The Price of Compression: The theory accounts for "compressed anchors." If you store quantized or projected versions of historical data to save space, the certification bound becomes looser, providing an explicit mathematical "price" for memory efficiency.

Key Results

In the UCI-HAR study, CABA demonstrated a clear, tunable trade-off between plasticity and retention.

  • Zero Forgetting: While unconstrained finetuning resulted in 1.01 percentage points of forgetting, CABA (at a budget of $\epsilon = 0.05$) reduced measured forgetting to 0.00.
  • Accuracy Trade-off: This stability came at the cost of final average accuracy, which dropped from 93.04% (Finetune) to 88.48% (CABA).
  • Comparison to "Frozen": Simply never updating the model (Frozen) also yields zero forgetting but results in much lower final accuracy (84.42%), showing that CABA successfully allows for some beneficial adaptation while protecting old knowledge.
Figure 1: Three-seed UCI-HAR results. Bars show means; Table 2 reports sample standard deviations. Accuracy uses the full 0–100% scale, and lower forgetting is better. CABA reaches zero measured forgetting but trades off final accuracy; Frozen shows why forgetting alone is insufficient.
Table 2: UCI-HAR continual adaptation. Mean ± sample standard deviation over three seeds. Higher accuracy and lower forgetting are preferred.

Limitations and Future Work

  • Coverage Assumptions: The "certified" part of CABA relies on the anchor set being a $\rho$-net that covers the distribution support. In high-dimensional spaces (like raw images or complex LLM prompts), ensuring this coverage is extremely difficult.
  • Lipschitz Constants: The bounds require knowing the Lipschitz constant ($L$) of the divergence. While the authors derive this for affine softmax heads, estimating it for deep, non-linear networks remains a challenge.
  • Scope of Experiment: The current proof-of-concept is limited to a linear head on a sensor dataset. The authors explicitly state that this does not yet establish performance for large-scale foundation models or complex robotic tasks.
  • Future Directions: The authors suggest future work should focus on applying CABA to LoRA updates for LLMs and Vision-Language-Action (VLA) models, where the "anchors" would be versioned conversation histories or observation-instruction pairs.

Summary updated

Abstract

Continual adaptation requires deciding not only how to update a model, but whether to commit an update without exceeding a declared budget for changes to protected behavior. We introduce Certified Anchor-Based Adaptation (CABA), a candidate-agnostic post-update gate that reserves historical observations from candidate optimization, compares old and candidate policies on frozen representations, and selects the largest tested interpolation coefficient whose worst-anchor forward KL divergence remains within budget. The finite check requires neither action labels nor complete trajectories. We develop a theory-led certification framework that separates measured anchor retention from conditional off-anchor guarantees. Under explicit representation-coverage and Lipschitz assumptions, an accepted update satisfies a uniform behavioral-divergence bound on protected support. The analysis gives regularity bounds for affine softmax heads, probabilistic coverage and compression guarantees, interpolation monotonicity conditions, grid-resolution bounds, and trajectory-level and sequential-composition guarantees. In a controlled UCI Human Activity Recognition study used as a non-stationary sensor-policy proxy, CABA reduces mean forgetting from 1.01 percent for unconstrained adaptation to 0.00 percent at budget epsilon = 0.05, while reducing final average accuracy from 93.04 percent to 88.48 percent. The experiment illustrates a retention-plasticity trade-off and does not establish foundation-model or robotic performance. CABA provides an auditable retention check for candidate generators exposing comparable policy distributions and a declared update path, while deployment also requires separate evidence of current-task utility.
BibTeX
@misc{ariq2026certified,
  author = {Ahanaf Hasan Ariq and Abrar Shahid},
  title = {{Certified Behavioral Retention for Continual Adaptation with Held-Out Anchors}},
  year = {2026},
  eprint = {2610.40740},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.40740},
}
Figure