Researchers at New York University and the University of Calgary report that Meta’s Llama 3 8B Instruct model, trained with reinforcement learning on the GSM8K maths benchmark, matched full fine-tuning by deleting fewer than 0.05% of its weights: 83.55% accuracy against 83.24%. Its adapter is 8.4 MB against a 30,633 MB fine-tuned checkpoint, more than 3,600 times smaller by our arithmetic. The paper, The Surprising Effectiveness of Deleting Weights in LLM Post-Training, went up on arXiv on 7 November and is listed for NeurIPS 2026.
First author Jack Lu announced the method, Deletion Tuning, on X on 10 November, calling it “a step toward continual learning”. All figures here are the authors’ reported results.
Business Pill · Fine-Tuning
A short explainer of fine-tuning: how a general AI model is adapted to one job. It teaches the general idea only and does not discuss this paper.
The key insight: As we read it, the paper moves the unit of model customization from a copy of the model to a list of weights to switch off: on Llama 8B the authors’ mask is 8.4 MB against a 30,633 MB fine-tuned checkpoint, with the same accuracy.
The method trains no new weight values. In the project page’s words, “Deletion Tuning learns which pretrained weights to delete and writes no new values”. The paper describes a trainable score for each weight in the gate projections of the transformer blocks. At deployment, it says, a weight either keeps its pretrained value or is set to zero, so the adapter is a binary mask of one bit per weight before compression.
The authors tested it on Llama 3 and 3.2 Instruct models from 1B to 8B parameters and Qwen models from 0.5B to 30B, under three kinds of post-training: reinforcement learning with verifiable rewards (GRPO), supervised fine-tuning and self-distillation. Under GRPO, Deletion Tuning matched full fine-tuning on Llama 8B (83.55% vs 83.24% on GSM8K), Qwen 7B (52.32% vs 51.89% on a six-benchmark maths mean), Qwen 14B (55.43% vs 54.87%) and the Qwen3-30B-A3B mixture-of-experts model (58.54% vs 58.43%).
At smaller sizes it recovered 90% to 99% of full fine-tuning’s accuracy, and the authors write: “These results suggest that deletion becomes more effective as models grow”.
Adapter size on disk for each method on Llama 3 8B trained with GRPO on GSM8K, log scale, with each method’s accuracy. The Deletion Tuning mask is 8.4 MB against 30,633 MB for the fp32 full fine-tuned checkpoint. Source: arXiv 2611.10866, Table 1; the ratio in the title is our arithmetic.
What Was Posted on X
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning’s accuracy at scale. It’s thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
Jack Lu, the paper’s first author, posted at 18:08 UTC on 10 November that zeroing under 0.1% of weights matches full fine-tuning’s accuracy at scale. The paper itself puts the deleted share below 0.05% in every run of its main reinforcement-learning table; the piece uses the paper’s figures and labels them as the authors’ results.
Learning a 30 MB binary mask is all it takes to match full fine-tuning (FFT) on a 61 GB, 30B mixture-of-experts model, by marking only the few weights to delete.
As models scale up, Deletion Tuning becomes as accurate as FFT under RL, from 7B to 30B, dense and MoE alike, even on held-out benchmarks.
The second post of Lu’s thread gives the size claim for the largest model tested: a 30 MB mask against a 61 GB, 30B mixture-of-experts model. The paper’s own table compares the 29.6 MB mask with a 116,471 MB fine-tuned checkpoint stored in fp32, which is the basis the piece uses.
Smaller Than a Rank-1 LoRA
The paper compares adapter sizes in bytes on disk. On Llama 8B the Deletion Tuning mask is 8.4 MB; a rank-1 LoRA adapter is 10.0 MB, a rank-4 LoRA 40.0 MB, and the full fine-tuned checkpoint, stored in fp32, 30,633 MB. On the 30B mixture-of-experts model the mask is 29.6 MB against 116,471 MB for full fine-tuning, about 3,935 times smaller by our arithmetic. The abstract says “the trained mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint”.
By our arithmetic, the disk space of one fine-tuned Llama 8B checkpoint would hold about 3,600 of these masks, and a rank-4 LoRA is 4.8 times the size of the mask. The project page describes the deployment step: “changing tasks loads only a small bitmask, applied to one shared copy of the base weights, rather than a separate full model checkpoint”. For background on the general technique, see our explainer on how fine-tuning adapts a general AI model to one job.
Base model accuracy against Deletion Tuning under GRPO for the four largest backbones, with the full fine-tuning score and mask size under each model. Source: arXiv 2611.10866, Tables 1 and 2.
Less Forgetting, the Continual-Learning Angle
The authors also tracked how much each method disturbed six earlier skills (HellaSwag, MMLU, TruthfulQA, WinoGrande, IFEval and HumanEval) during maths training on the three Llama sizes. Full fine-tuning ended 1.3 to 3.1 points below the base model on that suite, and LoRA and random sparse fine-tuning fell up to 5.4 points below it. For Deletion Tuning the paper reports that “its prior-task accuracy stays within one point of base throughout training”.
On repeated sampling, Deletion Tuning at 128 attempts solved 99.62% of GSM8K problems on Llama 8B against 95.29% for the base model, which the authors read as new capability rather than only sharpening. They say “a natural next test is continual learning over a sequence of tasks”, and that masks for different tasks might be combined to merge skills without further training.
The Structural Read
As we read the size table, the comparison that matters for serving is with LoRA, the adapter most teams already ship: the mask is smaller than a rank-1 LoRA at every scale the paper tests, and a rank-4 LoRA is 4.8 times its size on Llama 8B by our arithmetic.
As we read the forgetting result, the authors measured what a skill costs the rest of the model: full fine-tuning ended 1.3 to 3.1 points below base on six earlier tasks, while the paper reports Deletion Tuning stayed within one point.
Lu, Yang et al., New York University and University of Calgary, arXiv 2611.10866
“post-training can work by selection alone”
Three Implications
FOR TEAMS SERVING MANY CUSTOM MODELS The project page says switching tasks loads only a small bitmask applied to one shared copy of the base weights; by our arithmetic one fine-tuned Llama 8B checkpoint holds the bytes of about 3,600 masks.
FOR ON-DEVICE AND EDGE DEPLOYMENT The adapter stores no floating-point values, only which pretrained weights to zero; the paper reports about 0.03 bits per target weight after compression.
FOR CONTINUAL-LEARNING RESEARCH The authors name continual learning over a sequence of tasks, and merging masks for different tasks, as next tests; neither is run in this paper.
The framework argues: “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.”
As we read it, Deletion Tuning puts a skill learned by reinforcement learning into a small mask over a public base model, so the asset worth owning becomes the reward signal and the mask, not the copy of the weights. The framework predates this paper and does not discuss it.
What Is Not Established
The paper does not test continual learning over a sequence of tasks or the merging of masks; the authors list both as future work. They say Deletion Tuning “uses less training memory than FFT but more than LoRA”; as we read it, against LoRA the advantage is in the stored and served adapter, not in training memory. Results cover maths, knowledge, reasoning, tool use and science benchmarks on Llama and Qwen models up to 30B parameters, and parity with full fine-tuning holds at the largest size tested in each setting, not at every size.
The paper’s full fine-tuning sizes assume fp32 checkpoints, so the size ratios are against fp32; Lu’s thread gives the 30B model as 61 GB. The author list names New York University and the University of Calgary; Lu’s X profile also lists Meta, and the paper does not say whether Meta uses the method. The size multiples here are our arithmetic from the paper’s tables. We did not contact the authors.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
A NeurIPS 2026 paper from NYU and the University of Calgary reports that deleting fewer than 0.05% of a model’s weights matched full fine-tuning on Llama 8B and on Qwen models up to 30B parameters. On Llama 8B the mask is 8.4 MB against a 30,633 MB checkpoint, more than 3,600 times smaller by our arithmetic, and the authors report prior-task accuracy held within one point of the base model.
A note on sourcing. We read the full paper, arXiv 2611.10866, submitted on 7 November 2026 by researchers at New York University and the University of Calgary and listed for NeurIPS 2026, together with its project page and first author Jack Lu’s nine-post thread on X of 10 November. All accuracy and size figures are the authors’ reported results; the size multiples are our arithmetic from the paper’s Tables 1 and 2. We did not contact the authors.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
