Language Model Scaling

Follow this topic Ask about this topic

Momentum

6 papers in the last four weeks, up 100% on the four weeks before. 0.1% of all new papers.

Jul 13Week of Sep 28

Latest papers 37

Oct 5, 2026cs.AI

Do Small Language Models Learn to Negotiate? A Controlled Scaling Study of RL-Trained Sellers

LLM agents are starting to own the full customer experience. Soon, LLMs may be selling and buying on behalf of companies and customers respectively. Small models are more cost-efficient at scale, but can reinforcement learning train them into competent sellers? We train four Gemma 4 checkpoints (2.3B to 31B effective parameters) with GRPO on a programmatic utility reward for bilateral multi-issue bargaining, and evaluate every arm on the same 1,152 negotiations against two frontier buyers it never saw in training. With the same learning rate (

10^{-6}

) for every size, the gain of the RL model over its base rises from

+0.001

at 2.3B to

+0.078

at 31B. Each size was trained once and the two smallest checkpoints use a different architecture, so we fit no scaling law. Tripling the learning rate, with the same or fewer training steps, improves on the shared rate at every size by

+0.032

(2.3B) to

+0.081

(4.5B). In exploratory comparisons with two frontier models run as sellers, the 12B seller trained at the tripled rate scores above both, though its untrained base already scores as high as they do. The 4.5B seller at that rate shows no detectable difference from either and fits on one 48 GB GPU. A further 2.3B arm at ten times the shared rate raises pooled score, but its gain concentrates on the evaluation buyer that shares a model family with the training pool. These results suggest tuning the learning rate before concluding that a small model cannot learn to negotiate, and testing against buyers from more than one model family.

Pedro Tabacof, Sagar Joglekar

Fin AI Research

RL for Language Models Multi-Agent Negotiation Small Language Models Language Model Scaling

Oct 5, 2026cs.CL

Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation

Tokenizer-free language models remove the inductive bias of fixed tokenizers by modeling text directly as bytes, but the resulting longer sequences substantially increase computation and eliminate explicit text abstractions. We ask whether this additional computation can be useful, and whether standard Transformers can learn the abstractions that tokenization provides. We study these questions on Transformers without specialized tokenization-related architectures. With token-superposition training and hash embeddings, byte Transformers consistently outperform subword Transformers as model size scales. We further find that byte Transformers build local text abstractions as external tokenizers: a set of segmentation-like positions are used to collect local context representations, and restricting up to

25\%

of intermediate layers to these local representations preserves downstream performance. Finally, these learned structures induce highly non-uniform generation difficulty, with uncertainty concentrated near local structure boundaries; exploiting them for speculative decoding yields

3.4\times

more accepted tokens than in subword Transformers.

Jie Wang, Shiwei Luo, Qi Zhang +1

School of Computer Science, East China Normal University, Shanghai, China · School of Computer Science, Fudan University, Shanghai, China

Language Model Scaling Byte-Level Language Model Language Modeling

Sep 30, 2026cs.LG

ID Balancing: Stable Training of Extremely Sparse MoE via PID-Based Load Control

Scaling Large Language Models (LLMs) via Mixture-of-Experts (MoE) enables massive parameter growth with nearly constant per-token computation. However, further scaling the parameter count requires increasingly sparse routing, where expert load imbalance becomes more severe. This imbalance reduces parameter utilization and training efficiency, and can undermine training stability, becoming a bottleneck to reliable scaling. In this work, we unify two representative auxiliary-loss-free methods as incomplete Proportional-Integral-Derivative (PID) controllers: DeepSeek's loss-free method acts as a fixed-step integral controller, while Kimi K3's Quantile Balancing functions as a generalized proportional controller. Building on this control perspective, we propose ID Balancing, an Integral-Derivative controller. It scales its integral term with load error and activates its derivative term only when imbalance worsens, enabling stronger corrections for large or worsening errors and smaller updates near balance. Evaluated across Top-

10

, Top-

5

, and Top-

3

routing over

768

experts, ID Balancing reduces worst-case backbone MaxVio and training-average backbone MinVio by over

50\%

and

12\%

, respectively, relative to the best baselines in the Top-

3

setting. When the total parameter count increases from

18.9

B to

69.9

B (Top-

10

-of-

768

), ID Balancing's worst-case backbone MaxVio remains nearly unchanged and is approximately

89.6\%

lower than that of the auxiliary-loss baseline. ID Balancing also maintains competitive language-modeling and downstream performance. The advantages of ID Balancing grow as sparsity increases, making it a promising solution for scaling larger, sparser MoE models.

Peng Jin, Zihan Qiu, Zekun Wang +8

Qwen Team, Alibaba Token Hub, Alibaba Group

Expert Load Balancing Mixture-of-Experts Language Models Language Model Scaling Feedback Control

Sep 29, 2026cs.LG

Trajectory Soup: Pushing the Compute-Scaling Frontier of LLM Mid-training via Diverse Trajectories

Mid-training equips pretrained large language models with specialized and reasoning capabilities, but the returns of this stage are bounded since additional serial compute yields little further downstream improvement and can even degrade some capabilities, which places a practical ceiling on how much compute mid-training absorbs. We revisit how this compute should be allocated to a single run or multiple similar optimizations. We find that branches forked from a shared checkpoint under various controlled recipe reaches measurably different regions of parameter space, and establish a form of compatible diversity that extending one run cannot supply. Therefore, we introduce Trajectory Soup, which distributes a mid-training budget over several independent branches, and consolidates strongest checkpoints selected on validation through intra- and inter-trajectory averaging into a single model. A local bias and variance analysis separates the two averaging levels, showing that inter-trajectory averaging removes residual error beyond the reach of averaging within a trajectory, while checkpoint selection carries a bias that bounds how many checkpoints are worth merging. Across model scales, learning-rate schedules, token budgets, and trajectory counts, Trajectory Soup improves aggregate downstream performance over the strongest single-trajectory average under matched budgets and keeps improving as budgets expand, with the advantage preserved after an identical post-training pipeline. These results position trajectory allocation and merging as a practical way to extend the compute-scaling frontier of mid-training beyond serial saturation.

Zhehao Huang, Changxin Tian, Qingyuan Yang +5

Ling Team, Ant Group · Shanghai Jiao Tong University

Language Model Scaling Model Merging LLM Training

Sep 28, 2026cs.AI

An Exact Generate - Transform Decomposition of Small-LLM Team Scaling Across Orchestration Architectures

Replacing one LLM agent with a collaborating team can raise accuracy, but whether scaling the team helps, and which architecture to scale, is unclear. Sweeping eight agent orchestration architectures across five instruction-tuned 7-9B models, five short-answer benchmarks, and an executable-code benchmark up to 30 calls, we find that the returns to team scaling are sharply task-dependent: from three to thirty calls accuracy rises by up to 17 points on the two arithmetic word-problem benchmarks (GSM8K, GSMHard) but by at most four on ARC, GPQA, and MMLU, for every architecture, a split the usual task-averaged number conceals. Proposer-Critic captures the arithmetic gains, scaling steepest and, in aggregate, surpassing every other architecture at the largest budget (item-clustered intervals exclude zero), though it ranks among the weakest elsewhere, and no architecture wins across tasks. We explain these trajectories with an exact generate-transform decomposition. Partitioning any workflow into proposal coverage and a downstream transform, any accuracy change splits exactly into an extensive coverage dividend and an intensive transformation change. The decomposition diagnoses each task: arithmetic offers coverage headroom that a critic-guided transform converts, whereas the multiple-choice benchmarks either saturate in coverage or fail to convert it, and on open-ended code generative recovery nearly vanishes so accuracy tracks coverage. At equal call budgets token cost still varies 2.1x. Extra calls therefore create candidate opportunity that only some architectures, on some tasks, convert. Team scaling is a task- and architecture-specific bet, not a uniform lever.

Blaz Bertalanic, Carolina Fortuna

Jožef Stefan Institute, Ljubljana, Slovenia

Multi-Agent LLM Systems LLM Agent Orchestration Language Model Scaling

Sep 28, 2026cs.CL

When Harness Beats Scale, and When Reading Beats Both

We describe our system for DocSem, the document-grounded quantitative reasoning shared task at DocInsights 2026, and analyze why it succeeded on labeled data and failed on the test set. The pipeline pairs hybrid block retrieval with Program-of-Thoughts (PoT) generation executed in a sandboxed interpreter, self-consistency sampling, and entity enrichment from chunk-level knowledge graphs. On our held-out split, application architecture moved the metrics far more than model scale did: PoT added 0.282 joint accuracy to a compact 7B model but at most 0.005 to a 72B model, and a 27B model with the full harness matched the 72B (0.884 vs.\ 0.873) at roughly 2.7

\times

fewer parameters and a quarter of the CO

_2

. We read this through a distinction between world knowledge, which scales steeply with parameters, and language knowledge, which scales gently, and show that structured-output training makes a compact model harness-ready rather than merely small. On the raster, watermarked test PDFs the same system collapsed to 13.58% joint (rank 149 of 163); a controlled re-rendering of the validation set reproduces the OCR half of the collapse while bounding what the simulation misses. Auditing the physical nature of evaluation inputs precedes architecture, and the leaderboard's bimodality is consistent with reading quality, not reasoning, having separated the field.

Ivan Bondarenko, Nikolay O. Nikitin

Novosibirsk State University · ITMO University

Numerical Reasoning in Language Models Language Model Scaling Document QA

Sep 24, 2026cs.CL

Likelihood Ranking doesn't Scale Like Prompting in LLMs

LLM evaluation is commonly performed either by prompting models to produce answers or by scoring candidate outputs with likelihood-based metrics. In multiple-choice QA, however, standard likelihood-based scoring is still conditioned on the question and answer set, and can therefore leverage the same task-conditioned answer-selection interface used in prompting. We study a complementary protocol based on likelihood ranking of declarative statements constructed from the same question--answer pairs. Across 95 decoder-only models, ranging from 0.1B to 104B parameters, and 10 MCQA datasets, we find a systematic divergence between declarative-statement likelihood ranking and prompted answering. Statement-likelihood accuracy remains comparatively stable across scale, whereas prompted answering improves sharply with scale and instruction-tuning. These results suggest that likelihood preferences over controlled declarative alternatives and task-conditioned answer selection probe distinct aspects of model behavior, and should not be treated as interchangeable.

Alessandro Bondielli, Lucia Passaro, Davide Bacciu +1

CoLingLab, Department of Philology, Literature and Linguistics, University of Pisa · Department of Computer Science, University of Pisa

LLM Evaluation Language Model Generation Evaluation Multiple-Choice Question Answering LLM Prompting Language Model Scaling

Sep 23, 2026cs.LG

KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling

Scaling a language model is not only a question of final quality: the architectural choice determines how much computation is spent during training, prompt processing, and autoregressive decoding to achieve certain model quality. An ideal model architecture should lower all above computation costs to facilitate scaling to a larger model, while ensure the larger model indeed outperforms smaller baselines. We introduce KV-Invariant Transformer Expansion (KITE), a scaling paradigm that achieves this goal. It trains the model from a smaller size to a larger size (i.e., saving training costs via upcycling), while places newly added parameters in regions that do not affect attention KV. Consequently, during inference, prefilling KV only relies on the smaller part of the model, so the inference costs are saved. As a concrete instantiation, we present Step Scale Transformer (SST), a two-tower decoder in which one tower produces KV and the other reads them. At comparable cumulative training compute, SST, a 67B MoE model with 2.15B active body parameters per decode token, achieves lower training loss than 47B and 63B MoE Transformers with 1.48B and 2.02B active body parameters, respectively, while reducing estimated inference cost by 6.7% and 31.6%.

Zhiheng Hu, Yixun Wei, Jian Zhou +8

StepFun

Language Model Scaling LLM Inference Acceleration

Sep 1, 2026cs.CL

Scaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs

How to divide a fixed annotation budget between supervised fine-tuning (SFT) and reinforcement learning (RL) during LLM post-training remains an open problem. Existing work characterizes only broad trends (e.g., SFT dominates in low-data regimes), lacks a principled allocation framework, and does not examine whether the optimal ratio transfers across model sizes. We frame this problem in terms of near-optimality: rather than seeking a single optimal SFT-RL ratio, we characterize the near-optimal region, the set of allocations within a specified tolerance of peak performance. Empirically, this region is wide even for small tolerances (2-10%), widens with model scale, and transfers reliably from small proxy models to large target models. This yields a practical strategy: small proxy-model experiments suffice to identify a transferable near-optimal region, eliminating the need for exhaustive large-scale search. Our results hold consistently across tasks, model families, and both preference-based off-policy and reward-supervision on-policy RL methods. We further analyze how the asymmetry in annotation costs between SFT and RL data shifts the near-optimal region.

Jingtan Wang, Arun Verma, Xiaoqiang Lin +4

Department of Computer Science, National University of Singapore, Republic of Singapore · Singapore-MIT Alliance for Research and Technology Centre, Republic of Singapore · Agency for Science, Technology, and Research (A*STAR), Republic of Singapore +1

Supervised Fine-Tuning Language Model Scaling LLM Post-Training RL Fine-Tuning

Aug 31, 2026cs.AI

When Does Bigger Help? A Controlled Study of LLM Scale for Ontology Learning

The effect of Large Language Model (LLM) scale on ontology learning (OL) performance remains insufficiently characterized. We present a controlled evaluation of 13 models spanning dense and Mixture-of-Experts variants from the Qwen3.5 and Qwen3.6 lineages, together with proprietary GPT release variants, using the OntoLearner retrieval-augmented generation pipeline. All models are evaluated with the same embedding model, retrieval configuration, prompt templates, decoding settings, datasets, and metrics on term typing, taxonomy discovery, and non-taxonomic relationship extraction across four biomedical and materials science and engineering ontologies. Within the dense Qwen3.5 lineage, increasing parameter count primarily improves precision rather than recall, with the largest gains occurring between 9B and 27B parameters. However, the effect of scale is neither monotonic nor uniform across tasks and domains. Dense 27B models outperform substantially larger sparse models on term typing, whereas larger Mixture-of-Experts models achieve the strongest open-weight results on taxonomy discovery. Non-taxonomic relationship extraction remains difficult across model scales, particularly for the Materials Data Science ontology. Performance differences across matched Qwen variants and proprietary GPT releases further indicate that architecture and model lineage can outweigh nominal parameter count. These findings show that model size alone is an insufficient selection criterion for OL and provide empirical guidance for reproducible LLM-assisted ontology engineering.

Hamed Babaei Giglou, Sören Auer, Jennifer D'Souza

TIB Leibniz Information Centre for Science and Technology, Hannover, Germany · L3S Research Center, Leibniz University of Hannover, Hannover, Germany

Language Model Scaling Ontology Learning

Aug 11, 2026cs.NI

Benchmarking LLM-Guided Control-Plane Policies for Backend Fault Isolation in HAProxy

Static load balancers cannot mitigate a backend that is degraded rather than down: round-robin and least-connections keep routing traffic to a server returning HTTP 500s until an operator intervenes. We ask whether a Large Language Model can replace the static routing policy itself, reading HAProxy and Prometheus telemetry every 10 seconds and isolating faulty servers through guardrailed calls to the HAProxy Data Plane API. On a reproducible benchmark with a persistent structural fault built into roughly one-third of a heterogeneous fleet, we sweep 15 open-weight models across five families (0.35B to 35B total parameters; dense, mixture-of-experts, and efficient-sparse architectures), reasoning modes, fleet scales of 3 to 9 backends, and two routing algorithms, totaling 240 runs. We find a capability threshold near 3B active parameters. Below it, LLM policies are typically unreliable and sometimes worse than no policy; above it, every model, regardless of architecture, saturates near an 88% reduction in client-perceived 5xx errors over the static baseline. The threshold is approximate: Gemma 4 E2B clears it with 2B active parameters, while the dense 3B Granite 4.0 Micro does not. The availability gain has costs. Draining concentrates load onto surviving servers, inflating tail latency 2.6 to 2.8 times, and enabling reasoning multiplies token spend roughly tenfold, overrunning the control interval and degrading effectiveness. The efficient operating point is a supra-threshold model in its cheapest non-reasoning mode, wrapped inside deterministic guardrails.

Aman Chauhan, Vishnu Pendyala

Language Model-Based Control Language Model Scaling LLM Routing

Aug 8, 2026cs.RO

Diminishing Returns of Intelligence: The Non-Linear Relationship Between LLM Scale and User Perception in Short-Duration Open-Ended Social Human-Robot Interactions

Large Language Models (LLMs) are increasingly used to drive embodied social agents, yet it remains unclear whether larger models improve user perception during brief human-robot encounters. This paper examines the effect of LLM parameter size on short-duration, open-ended social interactions with a robot interface. In a within-subjects study, 19 participants interacted with robot faces driven by Qwen3-VL models at 4B, 8B, and 30B parameters. Participants evaluated the interactions in terms of perceived intelligence, naturalness, enjoyment, and humor. Results showed no significant overall preference for the 30B model over the smaller variants, including no significant advantage over the 4B model in perceived naturalness or intelligence. A significant relationship between AI interaction frequency and intelligence rankings for the 30B model suggests that more experienced users may be more sensitive to differences in model capability. Overall, the findings indicate diminishing returns from model scaling in brief open-ended social HRI, where conversational flow, responsiveness, and socially appropriate behavior potentially matter as much as raw parameter count.

Amanda Rasille Røn Volf, Morten Roed Frederiksen

Computer Science Department of The IT-University of Copenhagen

LLM Evaluation Language Model Scaling Human-Robot Interaction

Aug 5, 2026cs.CL

Kathleen Writes: Autoregressive Generation and Data Scaling Without Attention

Papers 1-2 of the Kathleen series showed that a byte-level, attention-free architecture built from a wavetable encoder and multi-scale reverberant state can match strong baselines on classification at ~450-700K parameters, without pretraining. We ask whether the same ingredients can generate. (1) Scaling: on byte-level language modeling (WikiText-103, raw UTF-8, no tokenizer), the reverberant model beats a parameter-matched transformer at every dataset scale measured (2-512 MB), e.g. 1.84 vs 2.04 bits/byte at 512 MB with ~0.5M parameters; the transformer needs more than 512 MB to match what the attention-free model learns from 32 MB. (2) Measurement: we introduce FORM DISTANCE, a non-parametric, gaming-resistant instrument for "reads like text": nine statistical axes of human text define a reference cloud, and five constructed fakes are all rejected. (3) Generation: decoding policy dominates architecture -- widening the sampler halves the same model's distance (3.17 to 1.52), and a retrieval-augmented decoding scheme takes the frozen model further (1.52 to 1.14) with no training step involved; the ablation attributes the gain to the sparse phrase dose itself, not the selection gate. The gain has a sharp boundary condition: the phrases must come from the model's own training corpus -- a 40x larger foreign library helps not at all, an effect the attention twin shares, consistent with in-context integration being a capability of scale. We also report four architectural additions that did not help, and a computed lexicon reaching 94% of a learned table's top-1 accuracy at one fifth of the parameters. Everything runs offline; all experiments are reproducible on a free Kaggle T4.

George Fountzoulas

Department of Computer Engineering & Informatics · Frederick University, Nicosia, Cyprus

Language Model Generation Evaluation Autoregressive Language Modeling Language Model Scaling Language Modeling Generative AI Evaluation

Jul 30, 2026cs.CL

Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Decoder-only language models entangle long-term memory and reasoning in a single parameter set, making it difficult to scale memory capacity independently. Memory Decoder introduces a parametric long-term memory module but only studies it at a relatively small scale. In this work, we present Memory Decoder at Scale, scaling memory models up to 6.9B parameters and pretraining them on 300B tokens. At this data scale, the combined cost of indexing and search makes a standard Faiss pipeline infeasible. We address this bottleneck with a distributed pipeline for Faiss indexing and retrieval, together with sparse, batch-wise loading of kNN distributions. Across model scales, we find that allocating more parameters to memory yields a better parameter-performance tradeoff than scaling the base model alone. On 17 benchmarks, pairing a 6.9B general memory with Pythia-410M raises its average score from 29.86 to 37.34, surpassing Pythia-12B (37.24) with 39% fewer total parameters. For Qwen3 Base models ranging from 0.6B to 14B, 1.7B domain memories improve the average score across the three domains by more than 9 points at every scale. Overall, our results demonstrate that independently scaling pretrained memory offers a more parameter efficient path to improving language model performance.

Rubin Wei, Jiaqi Cao, Jiarui Wang +4

LUMIA Lab, School of Artificial Intelligence, Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Electronic Engineering, Tsinghua University

Language Model Pretraining Memory-Augmented Language Models Persistent Memory for Language Models Language Model Scaling Language Modeling

Jul 9, 2026cs.CL

Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models

Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining. We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed. Depth-recurrent (looped) Transformers pursue this goal but are hard to scale, because looped computation does not fit naturally with the pipeline parallelism used to train the largest models. We add computation along the sequence-length dimension, where the extra computation is simply a longer input and stays compatible with standard large-model training. We propose Hidden Decoding, a sequence-length scaling method applied during continued pretraining (CPT). It expands each token into n streams with independent embedding tables and keeps the intermediate streams' key-value cache as context, so each token performs more internal computation without adding or widening Transformer layers. To keep this affordable at scale, we introduce Stream-Factorized Attention, in which most layers attend only within each stream and only a few layers mix across streams, reducing the attention cost from quadratic to roughly linear in n. Experiments support two scaling results. At frontier scale, we train WeLM-HD4-80B and WeLM-HD4-617B at n=4 and improve their matched non-HD baselines, making Hidden Decoding the first demonstrated sequence-length scaling method at the 100B+ MoE scale. Across expansion factors, the gains grow as n increases, showing that sequence-length expansion is a practical fixed-backbone scaling path for frontier-scale LLMs.

Aiwei Liu, Cheng Shi, Chuhan Wu +45

Language Model Scaling Laws Transformer Attention Language Model Scaling LLM Training

Jul 9, 2026cs.LG

Understanding Layer Patching in Model Size Interpolation

Zero-shot model size interpolation aims to create new models of intermediate target sizes by combining existing models without additional training. Recent work on boomerang distillation [Kangaslahti et al., 2026] shows that a student language model distilled from a larger teacher can be expanded by iteratively patching its layers, replacing student layers with contiguous blocks of teacher layers to obtain models whose size and performance interpolate between the student and the teacher. In this work, we provide the first systematic study of student-layer selection for model size interpolation. We cast finding the optimal layer subset for each model size as an optimization problem and prove it can be viewed as a shortest-path problem in a certain acyclic graph. In experiments, we show that patching strongly shapes interpolation behavior, with effects that vary substantially across model families. We find that simple sequential strategies--patching either from the first layer to the last or from the last to the first--often achieve surprisingly strong performance in practice. We further introduce KLPatch, a greedy patching algorithm based on KL divergence, which often improves over last-to-first patching and approximately solves the optimization problem. Together, our results provide a principled understanding of how layer patching affects model size interpolation and offer practical guidance for constructing near-optimal interpolated models.

Sara Kangaslahti, Jonathan Geuter, Nihal V. Nayak +3

1Harvard University · 2Kempner Institute · 3IST Austria

Language Model Scaling

Jul 8, 2026cs.CL

Think Big, Search Small: Where Capacity Matters in Hierarchical Search Agents?

Large language model based search agents increasingly adopt multi-agent architectures in which a main agent decomposes a complex question into sub-queries and dispatches them to parallel sub-agents. However, existing systems instantiate all roles from a single model of identical scale, leaving open how model capacity should be distributed across roles. We factorize hierarchical search into three roles: a delegation role responsible for task decomposition, an execution role responsible for retrieval and evidence extraction, and an answer generation role held fixed as a confound control. We then conduct controlled capacity sweeps along the delegation and execution axes on five multi-hop QA benchmarks. The experiments yield three findings. First, role factorization consistently outperforms a single-agent baseline, improving exact match from 4.5 to 8.6 points across six model scales. Second, capacity sensitivity is asymmetric: scaling the delegation backbone improves EM by ~11 points, whereas scaling the execution sub-agent moves EM by only ~2.6 points, identifying decomposition as the capability bottleneck. Third, a 1.7B-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent in accuracy while consuming 37% fewer sub-agent tokens, advancing the Pareto frontier. These results suggest a concrete recipe for building hierarchical search agents: concentrate capacity at delegation and downsize execution without sacrificing accuracy. Our code is available at https://github.com/QinnanCai0115/role-factorized-search.

Qinnan Cai, Yibo Zhao, Xiang Li

School of Data Science and Engineering East China Normal University

Multi-Hop QA Multi-Agent LLM Systems Agentic Search Language Model Scaling

Jun 20, 2026cs.CL

Fixed RAG Compression Collapses Measured Reader Scaling

Retrieval-Augmented Generation (RAG) compression papers often evaluate a compressor on one to three readers and treat the compressed evidence layer as evaluation-neutral. We show this assumption is false: fixed compression can raise average accuracy while hiding reader upgrades and reversing model rankings. Across 20 readers and ten domain-method settings over four QA benchmarks and one summarization benchmark, compression gain decreases with reader baseline (nine of ten settings significant, p < 0.05). Generic summarization flips 31% of pairwise model rankings on LongMemEval-S, and a fixed HotpotQA compressor hides 80% of the raw upgrade from Qwen 7B to GPT-4.1-mini. Two opposing forces explain this paradox: compression rescues weak readers by removing noise they cannot filter, and harms strong readers by dropping details they would have used. The pattern appears across structured compilation, generic summarization, three trained compressor families, query-focused summarization, and an external audit of nine published compression papers. We release ragscale, a toolkit built on 177,000 row-level compression transitions, so any compression paper can audit reader scaling with three readers in one day.

Sugam Panthi, Rabab Abdelfattah

The University of Southern Mississippi

Retrieval-Augmented Generation LLM Evaluation Language Model Scaling RAG Evaluation Context Compression

Jun 16, 2026cs.CL

Variable-Width Transformers

Scaling model size, specifically depth and width, has driven significant progress in transformer-based language models. However, most architectures maintain a constant width across all layers, allocating a fixed parameter and computation budget evenly despite different layers potentially playing distinct computational roles. In this work, we empirically investigate nonuniform capacity allocation across network depth by proposing a

\times

-shaped > <former architecture. This design maintains wider early and late layers while narrowing the middle layers, utilizing a parameter-free residual resizing mechanism. Across decoder-only language models ranging from 200M to 2B parameters (dense) and 3B parameters (MoE), our > <former consistently outperforms parameter-matched uniform baselines on language modeling loss. By reducing the average layer width, this architecture also requires fewer overall FLOPs (22% reduction under fitted loss-matched scaling curves) and smaller KV cache memory and I/O cost (15% reduction). In analysis, we show that this bottleneck structure results in qualitatively different representations in residual streams. Overall, our results demonstrate that nonuniform width allocation can result in more resource-optimal scaling of language models.

Zhaofeng Wu, Oliver Sieberling, Shawn Tan +3

ãMIT · èMIT-IBM Watson AI Lab

Transformer Decoder-Only Language Models Language Model Scaling

Jun 5, 2026cs.LG

Reversible Foundations: Training a 120B Sparse MoE through State-Preserving Scaling

This paper reports on training a hundred-billion-parameter sparse mixture of experts on a single eight-GPU node, end to end. LightningLM 0.1V is a recurrence-backbone language model family grown in four stages from a small dense seed, through a 5B and a 9B mixture of experts, to a 120B model with 460 routed experts under top-12 routing. Each larger model is grown from the trained weights of the smaller one; active parameters rise monotonically from 1.78B at the dense seed to 5.93B at 120B (about 5% of the 118.67B stored). The full lineage runs on single nodes, the larger stages at 8K context, reaching a released training loss of 1.78 at 120B scale. This is a systems and experience report. It is organized around three disciplines. Reversibility: a reversible recurrence stack reconstructs activations in the backward pass instead of storing them, holding activation memory flat as the model grows. State-preserving growth: each expansion (dense to MoE, shallow to deep, few experts to many) is given as a reproducible principle paired with the failure that results from getting it wrong; several failures are silent. Single-node economics: the 120B trains through TQP, a strategy of quantized base expert weights and trained low-rank adapters that carries optimizer state on 2.26B adapter parameters rather than 100B+ resident in routed experts, cutting expert-path optimizer state by a factor of ~45. What is new is the integration of known primitives, not any primitive in isolation: one grown lineage running end to end on a single node, documented at practitioner level, with per-domain held-out loss as evidence that targeted capabilities (multilingual Indic competence, code) were learned by construction. Model family, tokenizer, and training code are released.

Rohan Shravan

The School of AI Bengaluru, India

Mixture-of-Experts Language Models Efficient Language Model Training Language Model Scaling LLM Training

May 28, 2026cs.CL

A Dual-Path Architecture for Scaling Compute and Capacity in LLMs

Looped transformers apply a shared block multiple times and have emerged as a parameter-efficient route to scaling compute in language models. However, at fixed FLOPs a looped model has strictly less capacity than a baseline transformer. We propose a novel dual-path block that can flexibly scale compute, the number of sequential operations applied to a hidden state, and capacity, the parameters available at a single step. For this we expose both axes as parallel pathways within a single layer: a deep sublayer re-applied K times with shared parameters, and a wide sublayer with an enlarged feed-forward network applied once. Independent per-token gates combine both axes and allow detailed per-token routing analyses. We show that across two FLOP budgets, our dual-path model surpasses iso-FLOP matched models on language modeling and downstream evaluations, while using fewer parameters than the baseline at matched FLOPs. The learned gates are directly interpretable and show systematic per-token allocation with function words and lexical content trend wide, while punctuation, symbols, and arithmetic tokens trend deep.

Markus Frey, Behzad Shomali, Joachim Koehler +1

Lamarr Institute · Fraunhofer IAIS · University of Bonn

Recurrent Transformers Language Model Scaling LLM Routing Adaptive Computation

May 20, 2026cs.CL

Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory

Scaling conditional memory offers a promising way to increase language-model capacity, but existing methods such as Engram learn large memory tables from scratch during pre-training, making memory scaling expensive and sometimes ineffective. We propose Memory Grafting, a conditional memory scaling method that utilizes frozen hidden states from a grafting model as conditional n-gram memory. Given frequent local n-grams, we run the grafting model offline, store final-token hidden representations as memory values, and let the recipient model retrieve them through exact longest-match suffix lookup. Retrieved memories are adapted by lightweight projections and gates, while a hash-based Engram fallback preserves coverage for unmatched contexts. Since the grafting model is only run offline and exact lookup has expected O(1) complexity with respect to memory-bank size, Memory Grafting expands external latent capacity with limited training and inference overhead. Experiments under matched recipient architectures and pre-training budgets show that Memory Grafting improves over both MoE and vanilla Engram baselines. In the 2.8B-scale setting, it improves the average benchmark score from 51.95 for MoE and 52.43 for vanilla Engram to 53.86. In the 0.92B-scale setting, all grafting-model variants improve over the baselines, with Qwen3.5-35B-A3B giving the strongest gains. These results suggest that pretrained models can serve as reusable constructors of external latent memory, providing a practical step toward scaling future language models beyond trainable parameters alone.

Runxi Cheng, Yuchen Guan, Yongxian Wei +7

Tsinghua University · Microsoft Reasearch Asia

Language Model Pretraining Memory-Augmented Language Models Language Model Scaling

May 19, 2026cs.LG

Introspective X Training: Feedback Conditioning Improves Scaling Across all LLM Training Stages

We tackle the question of how to scale more efficiently across the many, ever-growing stages of current LLM training pipelines. Our guiding intuition stems from the fact that the dynamics of later stages of the pipeline, e.g. post-training, can be used to inform earlier stages such as pre-training. To this end, we propose Introspective Training (or IXT), inspired by offline reward-conditioned reinforcement learning and applicable to any stage of training. IXT uses a thinking reward model to annotate data with natural language critique based feedback, enabling quality aware training from the earliest stages of the pipeline. Models are then trained by prefix-conditioning the data with the generated feedback -- ensuring that not all tokens are treated equally starting much earlier in training than usual. Comprehensive experiments on 7.5-12B transformer-based dense LLMs trained from scratch all the way up to 18 Trillion tokens seen show that our method: bends scaling curves resulting in up to 2.8x more compute efficiency generally; and reaches performance levels unachievable for models trained otherwise in domains such as math and code.

Brandon Cui, Ximing Lu, Jaehun Jung +7

1NVIDIA · University of Washington · 3Carnegie Mellon University +1

Language Model Pretraining Efficient Language Model Training Language Model Scaling

May 16, 2026cs.LG

Scale Determines Whether Language Models Organize Representation Geometry for Prediction

In language models, what a representation encodes is determined by the geometry of its representation space: distances, not activations, carry meaning. Existing tools characterize the shape of this geometry but do not ask what that shape is organized for. We introduce Subspace PGA, a metric that tests whether a layer's distance structure aligns with the readout subspace of the unembedding matrix

W_U

more than with random subspaces of equal size. Across seven Pythia models (70M--6.9B) and three cross-family models, intermediate geometry is significantly organized for prediction (peak

z = 9

--

24

), but the degree is scale-dependent: small models (

d \leq 1024

) progressively lose it at late layers during training -- even as loss keeps improving -- while large models (

d \geq 2048

) preserve it throughout. We trace this to a capacity trade-off: a few dominant directions migrate away from

W_U

's readout, masking rather than destroying the predictive structure beneath, and removing them restores alignment. Neither spectral metrics nor loss curves capture this distinction. Scale thus determines not only how well a model predicts, but how its representation geometry is organized to do so.

Weilun Xu

School of Computer and Communication Sciences ´Ecole Polytechnique F´ed´erale de Lausanne

Language Model Pretraining Neural Representation Geometry Language Model Scaling Representation Geometry in Language Models

May 13, 2026cs.LG

When is Warmstarting Effective for Scaling Language Models?

Model growth from a given checkpoint aims to accelerate training of a larger model, offering potential resource savings. Despite recent interest, warmstarting has seen limited practical adoption in large-scale training. We attribute this to two underexplored factors: (1) an overemphasis on preserving the smaller model's performance at initialization, which constrains operator design for new architectures, and (2) insufficient analysis of how growth interacts with hyperparameters and scaling behavior, compounded by inconsistent growth factors across the literature. We show that preserving the base model's initial post-growth performance is not necessary for strong final performance, and that simple, architecture-agnostic growth strategies can outperform more complex warmstarting operators. Crucially, we empirically identify an upper bound on the growth factor

g

beyond which training from scratch is more efficient. We observe this across multiple ablation setups. Notably, this limit is also present, but unreported, in prior published results. Across our experiments on dense MLPs and dense language models, we find that a

2\times

growth factor is the most reliable in yielding convergence speedups, with gains most pronounced under 20 tokens/parameter budgets and diminishing as budget increases. We fit scaling laws over these observations to provide predictive guidance for practitioners deciding when and how much to grow. Together, our analysis provides practical guidelines and empirical limits for model growth.

Neeratyoy Mallik, Maciej Janowski, Johannes Hog +4

University of Freiburg · Zuse School ELIZA · University of Technology Nuremberg +3

Language Model Scaling Laws Language Model Scaling

May 12, 2026cs.LG

Procedural-skill SFT across capacity tiers: A W-Shaped pre-SFT Trajectory and Regime-Asymmetric Mechanism on 0.8B-4B Qwen3.5 Models

We measure procedural-skill SFT contribution across three Qwen3.5 dense scales (0.8B, 2B, 4B) on a 200-task / 40-skill holdout, with Claude Haiku 4.5 as a frontier reference. The corpus is 353 rows of (task + procedural-skill block, Opus chain-of-thought, judge-pass) demonstrations. Main finding. Under matched-path LLM-only scoring, the SFT-attributable procedural-

Δ

lift is roughly uniform across sizes:

+0.070 / +0.040 / +0.075

at 0.8B / 2B / 4B. Variation in post-SFT

Δ

(

-0.005

,

+0.100

,

+0.065

) is dominated by a W-shaped pre-SFT base trajectory (

-0.075

,

+0.060

,

-0.010

, Haiku-4-5 at

+0.030

): the 5-step procedure hurts 0.8B and 4B, helps 2B, and helps frontier Haiku modestly. SFT works hardest in absolute terms where the base struggles with the procedure -- a regime-asymmetric pattern with a falsifiable prediction at 8B/14B. Methodology. (i) A bench format-compliance artifact: 83.5% of the holdout uses a deterministic ANSWER-line extractor that under-counts free-form-prose conclusions; our LLM-only re-judge reveals it was systematically biased against the curated condition. (ii) A negative-iteration sequence at 0.8B: three well-formed recipe variants cluster post-SFT curated pass-rate within a 2 pp band, constraining the absolute-pass-rate ceiling to base capacity rather than recipe. Cross-family judge validation. GPT-5.4 via OpenRouter on all 7 configurations (2800 paired episodes) agrees on the direction of every per-student finding: Cohen's

κ\geq 0.754

, agreement

\geq 93.25\%

, max headline

Δ

shift

\leq 0.035

pp. Two earlier framings -- "format-only learning at 0.8B" and "SFT contribution shrinks at 4B" -- were path-mismatch artifacts; this paper supersedes both. Single-seed evaluation; threats itemised in the paper.

Igor Strozzi

Applied Mathematics Department Federal University of Rio de Janeiro

Supervised Fine-Tuning LLM Evaluation Language Model Scaling Language Model Post-Training

May 9, 2026cs.LG

Sparse Layers are Critical to Scaling Looped Language Models

Looped language models repeat a set of transformer layers through depth, reducing memory costs and providing natural early-exit points at loop boundaries. However, looped models do not scale as favorably as standard transformers with unique layers. We compare standard and Mixture-of-Experts (MoE) transformers, with and without looping, and find two main results. First, we find Looped-MoE models scale better than the standard baseline while dense looped models do not. We trace this to routing divergence between loops: in Looped-MoE models, different experts are activated on each pass through the same shared layers, recovering expressivity without additional parameters. Our second finding is that looped models have better compute-quality trade-offs with early exits than standard models. Because each loop ends with the same layers that produce the final output, loop boundaries are superior exit points, as confirmed by earlier output convergence at these points. In sum, we provide a clear direction for scaling looped models: a Looped-MoE model with early exits can not only beat standard transformers at scale, but also enable significant memory and inference savings with minimal degradation in quality.

Ryan Lee, Jacob Biloki, Edward J. Hu +1

1USC Information Sciences Institute · 2Netflix · 3Independent Researcher

Mixture-of-Experts Language Models Recurrent Transformers Language Model Scaling LLM Inference Acceleration Language Modeling Recursive Language Models

May 8, 2026cs.LG

Scaling Categorical Flow Maps

Continuous diffusion and flow matching models could represent a powerful alternative to autoregressive approaches for language modelling (LM), as they unlock a host of advantages currently reserved for continuous modalities, including accelerated sampling and tilting. Recently, several works have demonstrated the possibility of generating discrete data continuously by a simple flow matching process between a Gaussian and the one-hot encoded data distribution. They have further shown the feasibility of accelerated sampling via Categorical Flow Maps (CFMs), resulting in competitive sample quality in the few-step regime. However, this method had only been evaluated at relatively modest scales (

<1

B), leaving the question of its scalability completely open. In this article, we train a

1.7

B-parameter base flow model on

2.1

T tokens and self-distill it into a CFM that generates diverse, high-quality text in as few as

4

inference steps while maintaining near-data-level token entropy. Furthermore, we introduce a likelihood bound for CFMs in the semi-discrete setting, and show that they can be used to score the model on standard LM benchmarks, achieving results in the same range as discrete diffusion methods. Finally, we uncover some of the challenges that arise from training these models at scale, and we provide prescriptive insights on loss weighting and time scheduling.

Oscar Davis, Anastasiia Filippova, Pierre Ablin +4

Apple · University of Oxford

Language Model Scaling Flow Map Learning Flow-Based Generative Modeling Continuous Diffusion Language Models Text Generation Discrete Flow Matching

May 1, 2026cs.IR

Seeking Information with RAG-Assistants: Does Model Size Matter in Human-AI Collaborations?

Much research on LLMs has focused on increasing benchmark performance. However, the evaluation of such models in real-world collaborative human-AI workflows has stayed behind. This work evaluates a chatbot-style assistant based on Retrieval-Augmented Generation (RAG) in a realistic multi-turn information-seeking scenario inspired by workplace settings where compliance with local legislation and secure handling of sensitive data are often key. Specifically, we examine the performance of humans (N=112) assisted by RAG-assistants compared to LLM-only or LLM+RAG baselines. In this setting, we investigate how underlying model size (3B, 8B, and 70B) shapes the human-AI collaborative dynamic and how it influences perceived usability and satisfaction. Results show that the performance gain of human-AI collaboration over the model-only baselines is significant, irrespective of model size, suggesting that hybrid systems are beneficial in information-seeking scenarios. Interestingly, however, perceived usability and satisfaction among participants showed little difference across model sizes. This demonstrates a nuanced trade-off between model size, performance, and user perception. Our work highlights the added value of evaluating AI applications in actual multi-turn interactions with human users, looking at usability and satisfaction besides accuracy, rather than focusing on benchmark performance only.

Lennard C. Froma, Tom Kouwenhoven, Maaike H. T. de Boer +2

Leiden University - The Netherlands · TNO - The Netherlands · TU Delft - The Netherlands

LLM Evaluation Language Model Scaling Information Retrieval RAG Evaluation Human-AI Collaboration

Apr 28, 2026cs.CL

Scaling Probabilistic Transformer via Efficient Cross-Scale Hyperparameter Transfer

Probabilistic Transformer (PT), a white-box probabilistic model for contextual word representation, has demonstrated substantial similarity to standard Transformers in both computational structure and downstream task performance on small models and small to medium sized datasets. However, PT is less robust to hyperparameter choices than standard Transformers, making it harder to scale efficiently. In this work, we follow Maximal Update Parametrization (muP) to rescale PT's parameters, so that hyperparameters optimized on small models can be transferred to larger models without additional tuning. With this approach, we successfully scale PT to models with up to 0.4B parameters. Experiments show that PT consistently outperforms standard transformer under the same parameter budget on Masked Language Modeling (MLM) tasks. We hope this work will contribute to the practical deployment of probabilistic models at substantially larger scales in the future.

Penghao Kuang, Haoyi Wu, Kewei Tu

School of Information Science and Technology, ShanghaiTech University

Transformer Neural Network Optimization Hyperparameter Transfer Language Model Scaling