Search arXiv⌕ Search

arXiv · 2610.11146

What Transfers from Training-Only Priors? Separating Knowledge, Ordinal Targets, and Optimization

Abstract

Training-only teacher access lets a compact student deploy with local features and exported weights. This established distillation interface does not identify what transfers or when it helps. We separate label, query, capacity, and source-access budgets on three public credit archives and a previously unexamined bankruptcy archive. Shared tree parameters, shuffles, and optimizer interventions expose limitations of historical Ranking Prior Alignment; no tested single repair consistently rescues it. Heldout source queries and matched warmup and mappings retain a rank-listwise advantage over mapped staged KD in 24 credit configuration means and all six teacher--capacity means of a frozen new-archive test, with explicit query failures and stronger source-weight delivery under a different interface. A pretrained TabICLv2 follow-up at 100 labels initially improves matched supervision in only one of eight means; six of nine source-predicted policy gains reverse on target. A registered seen-archive intervention adds 500 or 2,000 unlabeled queries: all eight matched AUC means become positive under selected and fixed alpha. However, outer effects are mixed and no conditional paired-draw interval for these supervised contrasts lies wholly above zero. Increasing teacher context at 500 student labels does not beat matched supervision in any of eight ordinary-context means, despite stronger direct teacher performance. Rank KD retains substantial probability errors. Independent reloads verify zero guarded external inference accesses. Results support query/context boundary diagnostics, without universal superiority, population significance, a prospective transfer rule, or an LLM-specific advantage.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Qiye Lu, Jiang Ji, Liang Zhang. 2026-10-23. What Transfers from Training-Only Priors? Separating Knowledge, Ordinal Targets, and Optimization. https://arxiv.org/abs/2610.11146

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

A Theoretical Bridge Between Long-Tailed Recognition and Continual Learning

We theoretically and empirically establish a previously unexplored connection between Long-Tailed Recognition (LTR) and Continual Learning (CL). Specifically, we show that training on a long-tailed dataset with imbalance factor $\mathrm{IF}$ drives model parameters into an $O(1/\sqrt{\mathrm{IF}})$ neighborhood of the dominant-class solution, under both uniform and exponentially decaying cardinality distributions. Building on this result, we establish an objective-level reformulation of LTR as a sequential learning problem and prove that the CL objective upper-bounds the balanced LTR loss up to a constant determined by the residual Head loss. Motivated by this connection, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled framework that leverages standard off-the-shelf CL methods to sequentially learn Head and Tail classes while mitigating catastrophic forgetting. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions and demonstrate strong performance across diverse LTR benchmarks. In the foundation-model regime, CLTR with DualPrompt on a CLIP backbone outperforms standard adaptation baselines and is competitive with specialized methods using external semantic supervision. Our work bridges LTR and CL both theoretically and empirically, providing a principled approach for addressing long-tailed learning using standard CL strategies.

cs.LG↗

TRAM: Test-Time Risk Adaptation with Mixture of Agents

Deployed reinforcement learning agents often face safety requirements that are specified only after training: new hazard maps, revised risk thresholds, or behavioral alignment constraints. We study zero-update deployment-time adaptation in which a fixed library of risk-neutral source policies must be reused under a newly specified reward-risk tradeoff. We propose TRAM (Test-time Risk Adaptation with Mixture of Agents), a source-scored composition rule that evaluates each source under the target reward and an occupancy-based deployment risk, then selects actions using risk-adjusted source scores. Unlike training-time risk-sensitive methods tied to a fixed surrogate such as return variance, TRAM supports spatial barrier exposure, divergence to a reference behavior, and local volatility risks specified at test time. We make the surrogate nature of the method explicit: TRAM is not claimed to solve the full occupancy-control problem of the stitched policy, but it admits a measurable source-hull mismatch term connecting source-scored risk to realized risk. Experiments in gridworlds, MuJoCo Reacher, Safety-Gymnasium, and an inference-time LLM alignment setting show that TRAM improves deployment risk while preserving reward and requires no parameter updates at test time.

cs.LG↗

Towards the Worst-case Robustness of Large Language Models

Recent studies have revealed the vulnerability of large language models to adversarial attacks, where adversaries craft specific inputs to induce incorrect or harmful outputs. Although various empirical defenses have been proposed, their worst-case robustness remains poorly understood, raising concerns about how they would behave against stronger future attacks. Here, we study the worst-case robustness of LLMs from both empirical and theoretical perspectives. First, we provide tighter empirical upper bounds for the tested deterministic defenses using stronger white-box attacks, under which most tested configurations achieve nearly 0\% robustness. Second, we formulate randomized-smoothing certification for bounded and binary base functions as fractional and 0-1 knapsack problems, respectively, and derive grouped solvers for two token kernels. For example, on AdvBench, we certify a GPT-4o safety detector smoothed by a uniform kernel to an average certified $\ell_0$ radius of 2.02. In our stated suffix-attack setting, the same kernel certifies a Llama-3-8B safety detector to an average certified radius of 6.41.

cs.LG↗