One deployed language model must often serve many compute budgets, yet serving each budget still means a separate training or compression run per point. We train a Telescopic Language Model (TLM) to be that continuum: a nested-capacity Transformer supervised by stochastic prefix supervision with a full anchor. At every step, one randomly truncated prefix of the capacity axis is trained against the full next-token target, alongside one full-capacity pass, so the trained artifact is a valid language model at every depth. Two forward-backward passes per step, no architectural change, nothing extra at inference. Fixed-exit suites such as Matryoshka Language Model Suites (MLMS) occupy one point in this design space, and the point has a cost: supervising only a few fixed exits leaves the nested model at chance level everywhere else (perplexity 10^2-10^5 in our baselines). On a 200M proxy suite (20B FineWeb-Edu tokens, identical data stream for all methods), a single TLM run is a valid language model at every one of its twenty layer prefixes, in perplexity and on perplexity-sensitive downstream tasks, reducing the area under the quality-budget curve by 43-44% relative to the fixed-exit suites while matching them at full capacity, at ~12% lower GPU cost per run. The prefix sampling density is a dial: concentrating it on a few depths recovers fixed-exit quality there at the price of the continuum, so the operating points become a training-time choice rather than an architectural one. These results indicate that the training objective, not the nesting itself, is what makes a model elastic.
Figures & tables
Method
200M PPL ↓
AULB ↓
LODAacc↑
LODAppl↑
MLMS (best suite)
14.98
5.90
21.8
27.8
TLM ( π uniform)
14.99
3.28
38.9
47.3
TLM ( π log-unif.)
14.99
3.45
34.9
48.5
TLM ( π exit-grid)
14.65
6.21
–
25.7
vanilla twin (ref.)
13.98
–
–
–
Table 1: Headline results (200M proxy suite, 20B tokens/run, identical data stream). Full-size (200M) validation PPL; AULB over the prefix grid k=1..20 ; LODAacc/ppl , the area under the quality–throughput frontier (§ 3.3 ; higher is better). Bold marks the best single-run cascade per column; the vanilla twin is the per-size reference. ‘–’: not evaluated.
Method
GPU-h ↓
operating points ↑
GPU-h / point ↓
vanilla suite (3 models)
216
3
72
MLMS suite, no distill (1 run)
149
3
50
MLMS suite, +distill (1 run)
182–185
3
61–62
TLM, uniform π (1 run)
131
20
6.5
Table 2: One run, twenty operating points: TLM delivers each at 6.5 GPU-h versus 50–72 for the suites. Training cost (billed GPU-hours, one A100 per run, 20B tokens/run); TLM and MLMS rows are single runs, the like-for-like row holds distillation off on both sides, and TLM’s twenty points include off-frontier prefixes (see text).
Modern language models, including transformer, recurrent, and memory-based variants, share a common chassis: a stack of identical layers in which parameters are allocated uniformly across depth. This is a default inherited from the original transformer and largely unchanged since, yet a growing body of evidence suggests that layers contribute non-uniformly to the final output, with later layers refining the residual stream rather than transforming it. We ask whether parameter capacity should reflect this asymmetry. Our controlled experiment shows that, under a fixed budget, allocating more capacity to earlier layers and less to later layers improves perplexity over a uniform-width baseline, while the reverse allocation hurts. Building on this result, we introduce Tapered Language Models (TLMs), an architectural principle in which a parameter-bearing component is monotonically tapered across depth under a fixed total budget. MLPs are the natural site for this instantiation: they dominate parameter count across all modern LM families and expose width as a single, clean axis of variation. Across three model scales and four architectures (Transformer, Gated Attention, Hope-attention, and Titans), tapering MLP width via a smooth cosine schedule consistently improves perplexity and downstream benchmark performance over uniform baselines, at no additional parameter or compute cost. These findings establish depth-aware capacity allocation as a simple, architecture-agnostic axis of language model design, a free lever hidden in plain sight.
Reza Bayat, Ali Behrouz, Aaron Courville
1Mila · 2Cornell University · 3Université de Montréal +1
Training a language model suite classically requires training each model separately and serving them independently. We improve both training and inference efficiency by stacking sub-models of increasing size into a single nested architecture trained end-to-end. This Matryoshka training framework reduces the total parameter count of the suite, enables low-cost distillation from the largest to all smaller sub-models at every training step, and is well-suited for speculative decoding as the draft model is contained within the verifier. We validate our approach by training a Matryoshka suite comprising 500M, 1.5B, and 3B sub-models. Our suite is on par with independently trained baselines on benchmark performance and validation and out-of-domain perplexities, while using 36% less training compute and improving the throughput of speculative decoding by 14-26%. We also ablate key architectural choices, offering guidance for building strong Matryoshka LM suites.
Training a family of large language models (LLMs), either from scratch or via iterative compression, is prohibitively expensive and inefficient, requiring separate training runs for each model in the family. In this paper, we introduce Star Elastic, a novel LLM post-training method that adds N nested submodels to a given parent reasoning model using the compute of one run (N-fold savings) via a single post-training job. Beyond reducing training costs, Star Elastic also addresses a fundamental limitation of efficient reasoning: the rigidity of static architectures, which forces the allocation of constant resources regardless of token difficulty. By unlocking elastic budget control, Star Elastic enables a novel inference scheme that uses different submodels for each reasoning phase (thinking and answering). Star Elastic supports (1) nesting along the SSM, embedding channel, MoE, and FFN axes, (2) learning nested submodels via an end-to-end trainable router, and (3) curriculum-based knowledge distillation. Building on the Nemotron Elastic framework, we apply Star Elastic to the NVIDIA Nemotron Nano models, with a particular focus on hybrid Mixture-of-Experts (MoE) architectures: from Nemotron Nano v3 (30B/3.6A), we generate 23B (2.8A) and 12B (2.0A) variants with 160B training tokens. All nested models match or outperform independently trained baselines of comparable size and achieve a 360x reduction versus pretraining from scratch and a 7x reduction over state-of-the-art compression. Crucially, elastic budget control advances the accuracy-latency Pareto frontier, achieving up to 16% higher accuracy and 1.9x lower latency via dynamic per-phase model selection. We further extend Star Elastic to quantized regimes via Quantization-Aware Distillation (QAD), producing nested NVFP4 and FP8 elastic checkpoints that preserve zero-shot slicing while delivering smaller deployment footprints.
Ali Taghibakhshi, Ruisi Cai, Saurav Muralidharan +17