SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.24124 · 2026

The Alignment Tax: Response Homogenization in Aligned LLMs and Its Implications for Uncertainty Estimation

Mingyi Liu

arXiv · PDF · Open in the Atlas

Code that ran

2 of 2 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 2 failures of our instrument; 0 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2603.24124.json · badge: /api/ran/2603.24124.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
DigitLion/ucbd-experiment community
Papers with Code archive
2 ran · 0 no recorded run
FunctionStatusWhere it lives
compute_knn_density Ran
ran; our contract draft was wrong, not the code
DigitLion/ucbd-experiment/run_v1b_density.py
pointer only (licence: NONE) · get_code("69952e1e19588bf2")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, DigitLion/ucbd-experiment when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

cosine_similarity Ran
ran; our contract draft was wrong, not the code
DigitLion/ucbd-experiment/run_v1b_density.py
pointer only (licence: NONE) · get_code("6b3330f03a270806")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, DigitLion/ucbd-experiment when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

Also present in other papers' code

Of 2 samples lifted from this paper's linked repositories, 2 carry a function identity and 0 do not. Of the 2 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 2 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

RLHF-aligned language models exhibit response homogenization: on TruthfulQA (n=790), 40% of questions produce a single semantic cluster across 10 i.i.d. samples under Jaccard bigram clustering (79% under embedding cosine clustering; robust across sample sizes N =3-10, temperatures T =0.3-1.5, and generation lengths 40-200 tokens; 33.5% SCR persists at 200 tokens on 200 questions vs. 0% for the base model). On affected questions, sampling-based uncertainty methods have zero discriminative power (AUROC=0.500), while free token entropy retains signal (0.603). This alignment tax is task-dependent: on GSM8K (n=500), token entropy achieves 0.724 (Cohen's d=0.81). A base-vs-instruct ablation on Qwen3-14B confirms the causal role of alignment: the base model shows 1.0% single-cluster rate vs. 28.5% for the instruct model (Wilcoxon p < 10 -6 ). A three-way training stage ablation (Base 0.0% → SFT 1.5% → DPO 4.0% SCR, all p < 0.003) localizes the cause to DPO, not SFT. Cross-family replication on four model families (Qwen3-14B: 28.5%, LLaMA-3.2-3B: 5.5%, Mistral-7B: 1.0% SCR) reveals alignment tax severity varies by family and scale. Cross-chain replication on Tulu-3 (Llama-3.1-8B→SFT→DPO+RLVR) shows minimal alignment tax (0.5% SCR vs. Zephyr's 4.0%), confirming severity is recipe-dependent. Stage-wise token entropy measurement (Base 1.175→SFT 1.055→DPO 0.776 nats, p=1.17×10 -40 ) reveals the mechanistic decoupling: DPO retains 66% of token-level entropy while collapsing response diversity, explaining why B1 retains signal where sampling-based methods fail. We validate this finding across twenty-five experiments, five benchmarks, four model families, and three model scales (3B-14B), with proxy (Jaccard), SINdex-style embedding (agglomerative cosine clustering), cross-encoder NLI-SE head-to-head (∆AUROC=+0.014 vs. Jaccard, CIs overlap), and canonical NLI-based baselines at three DeBERTa scales (large 435M, base 184M, xsmall 70M-all ≈0.51 AUROC). Cross-embedder validation with two independent embedding families (Qwen3-Embedding 78% SCR vs. Nomic-embed-text 92% SCR at τ =0.85) rules out coupling bias. Cross-dataset validation on WebQuestions (58.0% SCR at τ =0.85) confirms the alignment tax generalizes beyond TruthfulQA. LLM-judge labels are validated against Truth-fulQA gold answer templates (κ=0.487). The central finding-response homogenization-is implementationindependent and label-free. Motivated by this diagnosis, we explore a cheapest-first cascade (UCBD) over orthogonal uncertainty signals. Selective prediction raises GSM8K accuracy from 84.4% to 93.2% at 50% coverage; weakly dependent boundaries (|r| ≤ 0.12) enable 57% cost savings.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2603.24124")
get_code_for_paper("2603.24124")
have("2603.24124")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2603.24124.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2603.24124.svg)](https://syntology.ai/paper/2603.24124)

Connect an agent — have() is free.