Skip to content
OpenTrain AIFor AI Companies

Omega-S: A Functional Resilience Index for LLM Fine-Tuning

Alberto AcedoPublished Aug 4, 2026
arXiv PDF
Researcher verdict
Context only
Use as context only
Benchmark evidence
Thin evidence
Verify before relying
Time to first repro
A few days
Plan setup time
Risk flags
2
Review before use

Abstract

Domain fit: AI-core · Core AI workload signals detected from paper context and implementation/artifact evidence.

Fine-tuning a large language model on new data degrades what it previously learned. We present Omega-S, a drop-in penalty computed from the weight matrix alone: it needs no previous-task data, no Fisher matrix and no stored copy of the old weights. It is three lines in an existing training loop and adds under 4% to the cost of a step. Retention. On Llama-3-8B with LoRA, fine-tuned from code to prose and measured by HumanEval over ten seeds, Omega-S retains more of the original capability than no regularisation on 9 of 10 seeds (0.173 -> 0.238 absolute pass@1; sign test one-sided p=0.011, Wilcoxon p=0.006), as a retention ratio, 62.9% -> 84.1%. It also beats tuned weight decay on 10 of 10 seeds (p=0.002) and tuned EWC on 8 of 10 (p=0.014), every arm re-measured in the same session. Mechanism, measured rather than asserted. Omega-S is topological by construction, its objective built from Tr(A^3), but we measured which of its four factors actually moves and three do not: their elasticity with respect to the weights is at or below 1e-4, against 9e-3 for the degree-variance term. As implemented, the composite reduces to a penalty on the variance of node degrees, which means row magnitude in square modules and directional alignment in non-square ones. We report this because a method whose name promises one thing and whose gradient does another should say so. We also enumerate the open design choices, including a contrast-preserving construction that does what it was designed to do and makes retention worse on all ten seeds. Repeating an identical configuration, same seed and same hardware, gives a standard deviation of 0.104 in retention ratio. We have not found this quantified for low-rank fine-tuning of language models, and it bounds every seed-paired comparison in this literature, ours included. Code, per-seed results and the full record of negative results are available.

Results and benchmarks

Freshness tier: hot
Fine-tuning a large language model on new data degrades what it previously learned.

Implementation

No direct implementation yet

Maintained implementation evidence is not confirmed for this paper yet.

Use the implementation status and reproduction sections for the current action plan.

Implementation evidence summary
Confidence: low

Recommendation evidence is currently too limited for a maintained-repo choice. Use Implementation Status and Reproduction Path for a practical baseline plan.

Reproduction risks
  • Estimate is based on paper-only reproduction flow

Reproduction readiness

Time to first repro: days
Last checked: Aug 21, 2026

No repo

No verified implementation available

  • No maintained repository has been identified for this paper. Check adjacent implementations or HF artifacts below.

Hardware requirements

  • Expect multi-day setup/compute for meaningful reproduction based on current guidance.

Hugging Face artifacts

No trustworthy direct or curated related Hugging Face artifacts were found yet. Use targeted searches to quickly locate candidate models, datasets, and demos.

Tip: start with models, then check datasets and spaces if you need evaluation data or demos.

Research context

Tasks

LoRA

Methods

LoRA / Parameter-efficient tuning

Domains

Natural Language Processing, Large Language Models

Evaluation and human feedback data

Open this paper in HFEPX to review benchmark signals, evaluation modes, and human-feedback protocol context.

Open in HFEPX
Explore similar papers

Jump to Paper2Code search queries derived from this paper's research context.