Fine-tune an LLM by deleting weights. Nothing else.

Slide 1 · EFFICIENT POST-TRAINING

Fine-tune an LLM by deleting weights. Nothing else.

FROZEN WEIGHTS · LEARN ONLY WHICH TO SET TO 0

✕ = weight zeroed by the learned mask

WEIGHTS ZEROED

<0.05%

of all weights, and it matches full fine-tuning at the largest scale tested

THE PAPER

The Surprising Effectiveness of Deleting Weights in LLM Post-Training

Lu, Yang, Lasby, Pote, Ioannou, Ren · NYU, University of Calgary · NeurIPS 2026

THE IDEA

The adapter is a bitmask: keep or zero, one bit per weight.

Slide 2 · THE QUESTION

Can a model learn new skills if the only edit allowed is deletion?

Post-training normally writes new numbers into the network. This paper asks how little has to change: not new values, just which pretrained weights to switch off.

FULL FINE-TUNING

Any weight can take any new value.

all params

LORA

+ ×

Frozen W plus a small trainable low-rank update B·A.

new values, low rank

DELETION TUNING

Frozen W. The only allowed edit is setting a weight to 0.

1 bit per weight

Why it's plausible: prior work cited in the paper finds RL fine-tuning naturally updates only 5 to 30% of parameters.

Slide 3 · THE METHOD · DELETION TUNING

Learn a keep-or-zero mask on the gate projection

1 Freeze

Every pretrained weight W stays fixed

2 Score

One trainable logit G per gate-projection weight

3 Soft mask

Train with W · σ(α·G); α ramps 10³ → 10⁴

4 Hard mask

Deploy W · 1[G > 0]: keep or zero

WHERE IT LIVES · MLP BLOCK

x gate_proj up_proj SiLU × down_proj

DETAILS

• Gradients flow through the soft mask; the hard mask is used for evaluation and RL rollouts

• Starts from a Wanda pruning prior (~0.18% of gate weights zeroed); no fixed sparsity target

• Soft and hard masks end within 0.08 pts of each other

THE ADAPTER

1 bit

per gate weight. Base weights are untouched.

Slide 4 · KEY RESULTS

Deletion alone keeps up with full fine-tuning

All claims from the abstract and Sections 1–5 of arXiv 2611.10866. Fewer than 0.05% of weights are deleted in every GRPO run.

EIGHT BACKBONES · 0.5B TO 30B · TWO FAMILIES

RL · GRPO

GSM8K: Llama 3.2 1B/3B, Llama 3 8B SimpleRL-Zoo maths: Qwen2.5 0.5B–14B, Qwen3-30B-A3B

SFT

Llama 3.2 1B/3B on 14 MMLU-Pro subjects and 23 BBEH subtasks

On-policy self- distillation

Qwen2.5-7B-Instruct on ToolAlpaca (tool use) and SciKnowEval Chemistry

Baselines: full fine-tuning (FFT), LoRA r = 1 and 4 on all seven projections, and random sparse fine- tuning.

PARITY WITH FFT

GRPO · SFT · SDFT

on par with FFT at the largest scale tested in each, up to a 30B mixture-of-experts model

BEYOND THE BASE MODEL

0 / 128

With GRPO, it solves some problems the base model failed in all 128 sampled attempts.

Slide 5 · STORAGE & FORGETTING

A skill you can ship as a few megabytes, with less forgetting

ADAPTER SIZE · LOG SCALE

Mask, Llama 8B (paper, compressed)

8.4 MB

Mask, Llama 1B (paper, compressed)

1.0 MB

Raw bitmask, 8B gate weights (est.)

≈235 MB

Full 8B checkpoint, bf16 (est.)

≈16 GB

"est." = my arithmetic, not the paper's: Llama 3 8B has ~1.88B gate weights (1 bit each) and ~8.03B parameters (2 bytes each). 16 GB ÷ 8.4 MB ≈ 1,900×.

FORGETTING UNDER RL

−5.4

points on prior tasks: the worst drop for FFT, LoRA and random sparse fine-tuning. Deletion Tuning stays near the base model.

WHY IT MIGHT WORK

Its weight change is the closest to FFT's on stable rank, spectrum drift and principal-weight overlap (vs LoRA and random sparse). It also keeps forward KL on the new task low, which earlier work links to less forgetting.

Slide 6 · TAKEAWAY

Much of post-training may be switching off the right few weights.

CRITICAL PERSPECTIVE

• The headline is parity at scale: the claim is for the largest models tested in each setting
• Cheap to store, not to train: one logit per gate weight (~1.9B for an 8B model)
• Unstructured <0.05% sparsity gives no inference speed-up; only gate projections are masked
• Why deletion matches FFT is described (geometry, KL), not yet explained; no code linked yet

IF IT HOLDS UP

One base model could carry many skills as megabyte bitmasks, swapped in by deleting weights, with less forgetting than today's fine- tunes.

REFERENCES

arxiv.org/abs/2611.10866

arxiv.org/html/2611.10866

agenticlearning.ai/deleting-weights