The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models
arXiv:2607.28639v1 Announce Type: new Abstract: We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias. On unambiguous tasks (BBQ-disambig), response-based distillation from a Gemma-2-9B teacher improves context-following: for the most biased baseline (SmolLM2-1.7B-Instruct), it cuts the context-overriding error rate from 44% to 24%. On ambiguous tasks (BBQ-ambig), the same distillation destroys per-item refusal calibration: 15% of items where the baseline correctly abstained instead receive stereotype answers, even when overall refusal rate is preserved. The pattern reproduces on a second student family (OLMo-2-1B-Instruct), with silence-loss of 8% and filled-silence accounting for 89% of new bias. Across the full 28-configuration grid, the magnitudes of silence-loss and filled-silence are uncorrelated (Spearman $\rho=0.19$, n.s.), indicating that the two effects arise from distinct mechanisms. Aggregate stereotype metrics (CrowS-Pairs, overall BBQ Stereotype Reliance Score) average over both effects and conceal the per-item harm. We trace the calibration loss to a data-side mechanism: an audit of four training corpora finds <0.5% refusal-as-answer-shape. Supervised fine-tuning (SFT) with refusal injection either breaks parsing or over-corrects into a trivial-refuser regime (refusal rate 99.8%, disambig accuracy 0.2%) that aggregate metrics would call perfectly calibrated. We propose Per-Condition Calibration Diagnosis (PCCD), a three-step protocol that evaluates refusal calibration, context-following, and capability preservation. PCCD catches both the asymmetric harm and the trivial-refuser failure mode that aggregate evaluations miss.
-->
[Submitted on 19 May 2026]
Title:The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models
View a PDF of the paper titled The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models, by Plawan Kumar Rath
View PDF HTML (experimental)
Abstract:We show that knowledge distillation in small instruction-tuned language models has asymmetric effects on bias. On unambiguous tasks (BBQ-disambig), response-based distillation from a Gemma-2-9B teacher improves context-following: for the most biased baseline (SmolLM2-1.7B-Instruct), it cuts the context-overriding error rate from 44% to 24%. On ambiguous tasks (BBQ-ambig), the same distillation destroys per-item refusal calibration: 15% of items where the baseline correctly abstained instead receive stereotype answers, even when overall refusal rate is preserved. The pattern reproduces on a second student family (OLMo-2-1B-Instruct), with silence-loss of 8% and filled-silence accounting for 89% of new bias. Across the full 28-configuration grid, the magnitudes of silence-loss and filled-silence are uncorrelated (Spearman $\rho=0.19$, n.s.), indicating that the two effects arise from distinct mechanisms.
Aggregate stereotype metrics (CrowS-Pairs, overall BBQ Stereotype Reliance Score) average over both effects and conceal the per-item harm. We trace the calibration loss to a data-side mechanism: an audit of four training corpora finds
new | recent | 2026-07
Change to browse by:
cs cs.AI cs.CY
References & Citations
NASA ADS
Google Scholar
Semantic Scholar
Loading...
Data provided by:
Bibliographic Tools
Bibliographic and Citation Tools
Bibliographic Explorer Toggle
Bibliographic Explorer (What is the Explorer?)
Connected Papers Toggle
Connected Papers (What is Connected Papers?)
Litmaps Toggle
Litmaps (What is Litmaps?)
scite.ai Toggle
scite Smart Citations (What are Smart Citations?)
Code, Data, Media
Code, Data and Media Associated with this Article
alphaXiv Toggle
alphaXiv (What is alphaXiv?)
Links to Code Toggle
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub Toggle
DagsHub (What is DagsHub?)
GotitPub Toggle
Gotit.pub (What is GotitPub?)
Huggingface Toggle
Hugging Face (What is Huggingface?)
ScienceCast Toggle
ScienceCast (What is ScienceCast?)
Demos
Demos
Replicate Toggle
Replicate (What is Replicate?)
Spaces Toggle
Hugging Face Spaces (What is Spaces?)
Spaces Toggle
TXYZ.AI (What is TXYZ.AI?)
Related Papers
Recommenders and Search Tools
Link to Influence Flower
Influence Flower (What are Influence Flowers?)
Core recommender toggle
CORE Recommender (What is CORE?)
Author
Venue
Institution
Topic
About arXivLabs
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)