Pre-training giant language fashions is dear sufficient that even modest effectivity enhancements can translate into significant value and time financial savings. Nous Analysis is releasing Token Superposition Coaching (TST), a way that considerably reduces pre-training wall-clock time at fastened compute with out touching the mannequin structure, optimizer, tokenizer, parallelism technique, or coaching knowledge.
On the 10B-A1B mixture-of-experts scale, TST reaches a decrease last coaching loss than a matched-FLOPs baseline whereas consuming 4,768 B200-GPU-hours versus the baseline’s 12,311 — roughly a 2.5x discount in whole pre-training time.


The Downside TST is Fixing
Trendy LLM pre-training is closely data-driven. Latest coaching regimes routinely overtrain nicely past compute-optimal estimates, and uncooked textual content throughput. How a lot knowledge a mannequin can course of per FLOP has turn out to be a key lever. Subword tokenizers like BPE already enhance throughput by compressing sequences; and the analysis suggests a lot of the BPE benefit over byte-level fashions comes merely from shorter sequences, which implies the mannequin sees extra textual content per unit of compute.
TST asks whether or not that throughput lever will be pulled additional throughout coaching, independently of the tokenizer and with out completely altering the mannequin.
How TST Works: Two Phases
TST modifies the usual pre-training loop in two sequential phases:
Section 1 — Superposition: For the primary r fraction of whole coaching steps (the paper finds r ∈ [0.2, 0.4] to be near optimum throughout examined scales), the mannequin doesn’t obtain particular person tokens. As an alternative, the enter sequence of size L is segmented into non-overlapping baggage of s contiguous tokens. Within the embedding layer, every bag is collapsed right into a single latent “s-token” by averaging the s token embeddings. The transformer then processes a sequence of size L/s.
Crucially, every TST step is stored equal-FLOPs to an ordinary coaching step by rising the info sequence size by s occasions throughout the superposition part. As a result of every latent place corresponds to s supply tokens, the mannequin ingests s occasions as a lot textual content per unit of compute — that is what drives the throughput acquire.
On the output facet, every latent place predicts the subsequent bag of s tokens slightly than a single subsequent token. The usual cross-entropy loss is changed with a multi-hot cross-entropy (MCE) loss, which assigns equal likelihood mass 1/s to every token within the goal bag. The MCE loss reduces to a easy imply of normal cross-entropy phrases over the s targets — it may be applied utilizing the prevailing fused CE kernels already current in any main pre-training library, with out writing a brand new kernel or including an auxiliary head.
Section 2 — Restoration: After the superposition part, coaching resumes from the saved checkpoint with normal next-token prediction for the remaining 1 - r steps. The TST code is totally eliminated at this boundary to keep away from any experimental contamination. A transient loss spike happens on the transition, sometimes between 1 and a pair of nats, which resolves inside just a few thousand steps. After that, the recovered mannequin crosses beneath the equal-FLOPs baseline and stays there.
The mannequin produced on the finish of Section 2 is architecturally an identical to at least one produced by typical pre-training, with the identical next-token prediction inference conduct.
What the Experiments Present
TST was validated at 4 scales: 270M and 600M dense (SmolLM2 shapes tailored to the Llama3 modeling code, with the Llama3-8B tokenizer and untied enter/output embeddings — which makes the 270M mannequin equal in measurement to SmolLM2-135M and the 600M to SmolLM2-360M), 3B dense (SmolLM3 form), and a 10B-A1B MoE within the Qwen3 household. Coaching used the DCLM dataset for the smaller runs and a 50/50 mixture of DCLM and FineWeb-Edu for the MoE run. All runs used AdamW with the Warmup-Steady-Decay studying price schedule and have been run in TorchTitan underneath FSDP parallelism, on 64 NVIDIA B200 GPUs for the bigger fashions and eight B200 GPUs for the smaller ones.
On the 3B scale with bag measurement s = 6 and step ratio r = 0.3, TST at 20,000 steps reaches a last lack of 2.676 — practically matching a 36,000-step baseline at 2.677 — whereas utilizing 247 B200-GPU-hours versus 443. The 20k-step TST run scores 62.4 on HellaSwag and 66.3 on ARC-Simple, versus 62.3 and 65.9 for the 36k baseline.
On the 10B-A1B MoE scale with s = 16 and r ≈ 0.25, the TST run processes 2T knowledge tokens and achieves a last lack of 2.236, beneath the baseline’s 2.252 after 1.05T tokens, whereas beating it on all 4 reported benchmarks: HellaSwag (71.2 vs. 70.1), ARC-Simple (74.2 vs. 73.8), ARC-Problem (47.3 vs. 46.3), and MMLU (39.0 vs. 37.4).
The analysis group presents three comparability views in opposition to the baseline — equal-FLOPs, equal-loss, and equal-data. Beneath equal-FLOPs and equal-loss circumstances, TST constantly wins. Beneath equal whole token consumption, the baseline wins, as a result of TST’s efficient compute price range per knowledge token is smaller. This is a vital boundary situation that determines the place TST applies.
Two Distinct Mechanisms
An ablation research isolates the input-side and output-side elements. Each independently outperform the baseline; combining them produces additional enchancment with out indicators of interference. The authors interpret this as proof that TST is 2 orthogonal mechanisms slightly than a single trick.
The output-side mechanism — next-bag-of-tokens prediction — is conceptually associated to multi-token prediction (MTP). Not like MTP, which provides ok unbiased prediction heads and additional parameters, TST retains a single output head and replaces solely the goal. This makes it the least costly member of a rising class of future-signal auxiliary targets. Not like MTP, it exhibits constant positive aspects throughout all examined scales together with small fashions the place MTP has been proven to degrade efficiency.
The input-side mechanism has no direct analog within the current pre-training literature. The analysis group provides two believable explanations: it might implicitly regularize the embedding geometry (since many random s-grams of tokens should stay linearly separable as soon as averaged), or it might act as a type of pre-pre-training, exposing the mannequin to a coarser model of the true knowledge earlier than fine-resolution language modeling begins.
A focused ablation straight assessments what occurs when illustration continuity is damaged. The analysis group runs a 3B TST experiment the place the enter embedding and output LM head are randomly re-initialized initially of Section 2. The outcome: last loss jumps to 2.938 — worse than each the TST run (2.676) and the usual baseline (2.808). The Section 1 TST steps contributed nothing to the ultimate mannequin. This confirms that shared representations throughout each phases usually are not incidental to TST’s success — they’re what makes it work.
Marktechpost’s Visible Explainer
Key Takeaways
- Nous Analysis’s Token Superposition Coaching (TST) cuts LLM pre-training time by as much as 2.5x at matched FLOPs — no structure, tokenizer, or optimizer adjustments required.
- Section 1 averages contiguous token embeddings into baggage and predicts the subsequent bag by way of multi-hot cross-entropy; Section 2 reverts to plain next-token prediction from the identical checkpoint.
- Validated at 270M, 600M, 3B dense, and 10B-A1B MoE — TST beats the baseline on loss and downstream evals (HellaSwag, ARC, MMLU) throughout all scales.
- Optimum hyperparameters: bag measurement s ∈ [3–8] for smaller fashions, step ratio r ∈ [0.2, 0.4]; shared embeddings throughout each phases are essential — re-initializing them makes TST worse than the baseline.
- Commerce-off: TST consumes extra uncooked knowledge tokens per compute price range — finest suited to compute-bound coaching; the output-only variant is the choice for data-bound settings.
Try the Paper and Undertaking. Additionally, be happy to comply with us on Twitter and don’t neglect to affix our 150k+ ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you may be a part of us on telegram as nicely.
Have to accomplice with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and so on.? Join with us
