Skip to content

Back to the catalog

Overcoming Prior Barriers: Supervised Fine-Tuning under Long-Tail Distribution
arxiv.orgpaper

Overcoming Prior Barriers: Supervised Fine-Tuning under Long-Tail Distribution

Introduces "prior barrier", a measure of how strongly a pretrained LLM supports competing concepts over a target concept, and shows prior barriers follow a long-tail distribution, so tail concepts need more SFT instructions than head concepts. It derives a predictive risk bound for SFT under long-tail prior barriers and proposes PASS, an adaptive instruction selection method that builds reference-derived concepts, estimates each instruction's distinguishing evidence, and allocates the selection budget toward under-covered concepts. PASS beats seven state-of-the-art instruction selection methods across four backbone-budget settings, and ablation shows adaptive allocation beats uniform allocation.

TL;DR Long-tail "prior barriers" explain uneven SFT difficulty across concepts; PASS adaptively selects instructions under a budget and outperforms seven selection baselines.
Open original ↗
Document

What it claims

The paper models SFT as Bayesian updating of a pretraining-induced prior π over "concepts". It gives a formula for how many instructions a concept needs, and that number grows with the concept's prior barrier:

  • Prior barrier: B(r) = sup over competing concepts θ (at distance ≥ r from the target θ⋆) of log π(θ)/π(θ⋆).
  • Head concept: B(r) ≤ 0. No competitor is better supported than the target.
  • Tail concept: B(r) > 0. Some competitor is better supported than the target.

PASS turns this into an instruction selector. It scores each candidate instruction per concept, then spends the budget on concepts that remain under-covered.

How it works

Theory (Section 3)

  • Distinguishing evidence. ε(r) is the minimum, over competitors outside the r-neighbourhood, of the expected KL(P_θ⋆(·|x) ‖ P_θ(·|x)) under the instruction distribution. The target is identifiable when ε(r) > 0.
  • Posterior concentration (Thm 1). With probability ≥ 1−η, once n ≥ n_sep, the posterior mass q⋆ₙ(A_r) left on competitors is at most min(1, (1/m(r))·exp(B(r) − nε(r)/8)). Here m(r) is the prior mass near the target.
  • Head–tail gap (Cor 1).
  • Head concepts need n ≥ (8/ε)·log(1/(m δ)).
  • Tail concepts need n ≥ (8/ε)·(B + log(1/(m δ))).
  • So a tail concept costs roughly 8B/ε extra instructions. Instructions that separate the target from its hardest competitor (high ε) shrink that cost.
  • Predictive risk (Thm 2). Excess log-loss Eₙ ≤ α(r) + M_r·q⋆ₙ(A_r). Here α(r) is the worst local discrepancy near the target and M_r bounds the KL discrepancy of competitors.

Method (Section 4)

Concept construction (M1):

  1. Encode each reference (x, y) with the pretrained model M₀ to get a representation z.
  2. Cluster these hierarchically. Keep splitting a cluster until the cosine similarity of its members to their centroid is concentrated enough.
  3. Each final cluster is one concept θ. Its normalised centroid p_θ is the desired representation.

Competitor construction (M1):

  1. Keep the same inputs but swap in wrong responses. These come from available incorrect alternatives plus incorrect outputs generated by M₀.
  2. Each competing concept k gets its own normalised centroid n_θk.

Scoring and allocation:

  1. Score each candidate instruction separately against every concept. The score is its similarity to p_θ versus its similarity to the closest n_θk. The exact formula is cut off in the fetched text.
  2. M2. Use M₀ to estimate each concept's barrier and existing pretrained support. Allocate instructions incrementally toward concepts that are still under-represented.
flowchart LR
  R[Reference set T] --> C[Cluster into concepts]
  R --> W[Wrong responses<br/>given + M0-generated]
  C --> P[Desired centroid p_θ]
  W --> N[Competitor centroids n_θk]
  P --> S[Per-concept evidence score]
  N --> S
  D[Candidate pool D] --> S
  S --> A[M2: allocate to<br/>under-covered concepts]
  A -->|budget left| A
  A --> O[Selected subset of size n]

Evidence

The fetched text stops partway through Section 4.1. It contains none of the following:

  • the experiments
  • which backbones or budgets were used
  • the names of the seven baselines
  • any benchmark scores

The only results it gives are the abstract-level claims, which the existing summary already covers.

How to use it

No code is available yet: "the source code will be released upon publication."

From the method description, adopting PASS requires:

  • a reference set T that represents the target behaviour;
  • wrong responses for the reference inputs, either already available or generated by M₀;
  • forward passes of the pretrained model over the reference set and the entire candidate pool.

Limits

  • The text is truncated. It is missing the scoring formula, M2's allocation rule and how barriers are estimated, all experiments, and the appendix assumptions behind Theorems 1–2.
  • The theory rests on interpretation. π is interpreted as a prior over concepts induced by pretraining. The bounds assume a Gibbs-posterior view of SFT. The fetched text does not show how B(r) is measured on a real model.
  • Concepts depend on the reference set. Concepts come from clustering embeddings of T. Coverage therefore hinges on how representative T is and on the coherence threshold, which the fetched text does not state.
  • Dated v1 preprint (8 Oct 2026). The released code may differ from the description.

Techniques named here

Related in training