acceptodds
Under review as a conference paper at ICLR 2027

Scaling LLM Knowledge Boundaries via Distribution-Optimized Synthesis

Abstract

Knowledge injection via synthetic data is crucial for enhancing Large Language Models (LLMs). However, current synthesis methods simply stop at preset token counts or fixed data ratios, lacking awareness of knowledge distribution. This results in some knowledge points being sparse while others are redundant, limiting LLM knowledge boundaries. We revisit knowledge injection from a distribution perspective and hypothesize that an optimal knowledge density range exists for maximizing knowledge boundary expansion. We propose **KDoS** (**K**nowledge **D**istribution-**o**ptimized **S**ynthesis), a framework that introduces knowledge density as a controllable variable to quantify the concentration–dispersion degree of knowledge in semantic space, and uses a three-stage feedback mechanism to drive synthesis, shifting from blind generation to distribution-optimized synthesis. We construct Wikipedia-based synthetic data with varying knowledge distributions and conduct experiments on models from 0.6B to 16B (Qwen, Ling, LLaMA) and data scales from 1B to 5B tokens. Our key findings are: (1) an optimal knowledge density range consistently maximizes boundary expansion; (2) this range is stable across backbones and scales; (3) KDoS outperforms representative synthesis baselines across six factual knowledge benchmarks. Our work offers a new perspective and practical framework for synthetic data-driven knowledge injection.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Large Language Model, Knowledge Boundary, Knowledge Injection

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_usftOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how the knowledge density of synthetic data affects LLM knowledge acquisition under a fixed training token budget. The authors define knowledge density to measure data concentration in embedding space and use it to select synthetic training data. Experiments on factual QA benchmarks across different model and data scales show that an appropriate knowledge density can improve knowledge injection.

Critical strengths

The paper addresses an important practical question: how to improve LLM knowledge acquisition under a fixed training token budget. It introduces an explicit measure of semantic concentration and uses it to guide synthetic data selection, providing a concrete way to control data distribution.

Critical weaknesses
  1. The main observation that highly concentrated training data leads to worse factual QA performance is reasonably expected. The paper does not sufficiently establish what additional insight or practical value its density measure provides beyond general data diversity and coverage considerations.

  2. The training procedure is insufficiently specified. The paper does not clearly state the training objective, the final input–output format, whether knowledge points and logic chains are included in training, or which tokens contribute to the loss. These omissions limit reproducibility and make the knowledge injection results harder to interpret.

Questions for authors

Could the authors provide the prompt templates and target format used for training and evaluation? Are the models explicitly instructed to produce only a short answer, and do the training targets contain any reasoning rationales? In particular, how does the gold-answer loss evaluation account for valid responses that include introductory text or explanations before the answer? This clarification is important for distinguishing factual knowledge acquisition from changes in response-format preferences.

Reviewer_McVaOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether the semantic concentration of synthetic QA data affects factual-QA performance under a fixed training-token budget. It proposes Knowledge Distribution-optimized Synthesis (KDoS), which extracts knowledge-point lists and knowledge logic chains from seed questions, generates new questions from semantically related examples, filters the candidates with two LLM judges, and accepts or rejects candidates according to a target density.

The density is computed from the total token count and the mean distance of the embedded knowledge-point lists from their centroid, using the volume of an nn-dimensional hypersphere. The experiments vary this quantity across several model sizes and token budgets. The authors report a range associated with lower evaluation loss. In the main method comparison, KDoS obtains the highest average accuracy among the reported baselines, although it is not the best method on every benchmark.

Critical strengths

1. A practical data-selection problem.

When candidate generation is less constrained than the training-token budget, some form of selection is necessary. The experiments specifically examine how concentrating that budget on similar knowledge differs from distributing it across a broader set of questions.

2. Evaluation beyond a single model configuration.

The density sweep covers multiple model and data scales, and the proposed pipeline combines question generation, quality filtering, and candidate selection. Using the density estimate during synthesis, rather than only reporting it after dataset construction, is a useful systems contribution.

Critical weaknesses

1.The terminology is broader than the evidence.

The experiments measure factual-QA accuracy and answer-token loss under a particular generation and embedding setup. This does not by itself establish a general “knowledge boundary.” Likewise, the proposed density is based on mean distance to one centroid; it does not characterize the full knowledge distribution. “Knowledge collapse” is also too strong for the observed increase in loss at high density because its mechanism is not identified.

2.The experiments do not isolate what density control adds.

KDoS changes the candidate-generation process, applies two-stage quality filtering, and changes the selection rule. The current ablation does not appear to include the key condition that retains the same generated candidates and quality filters while removing density-aware selection. Comparisons with semantic deduplication, farthest-first selection, facility location, or cluster-balanced sampling on the same candidate pool would clarify whether the gain is specific to the proposed measure or follows from ordinary duplicate removal and coverage improvement.

3.The hypersphere formulation needs justification.

For fixed token count and embedding dimension, the proposed density is a monotonic transformation of the mean radius. It consequently ranks datasets in the same order as the radius itself. The same mean radius can arise from broad multimodal coverage or from a compact dataset containing a few distant outliers. This paper should show either that the hypersphere transformation adds practical value or that performance is stable across datasets with similar density but different cluster structure. Sensitivity to the embedding model is also not reported.

Questions for authors
  1. If the candidate pool, quality filters, accepted token count, and training procedure are held fixed, how does density-aware selection compare with semantic deduplication, farthest-first selection, facility location, and cluster-balanced sampling?

  2. Can report an ablation that retains candidate generation and quality filtering but removes only density-aware selection? It would also help to report the number of unique knowledge clusters and the distribution of fact-level repetition in each density condition.

  3. Since the proposed density is a monotonic transformation of the mean radius for fixed TT and nn, what is gained by the hypersphere-volume formulation? Are the dataset ordering and observed loss trend preserved with another embedding model?

Reviewer_P1f1OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KDoS, a distribution-optimized synthesis framework for factual knowledge injection into large language models. It defines knowledge density from the token count and the mean radius of knowledge-point embeddings, constructs a large Wikipedia-derived seed pool, generates and filters candidate questions, and applies density-aware rejection sampling toward a target distribution. Experiments span six factual QA benchmarks, several model families, model sizes from 0.6B to 16B, and data scales from 1B to 5B tokens. The paper reports an empirically effective density range and strong average performance relative to the evaluated synthesis baselines.

Critical strengths

S1: Extensive Empirical Study with Strong and Potentially Insightful Results. The paper involves substantial experimental effort, covering six factual knowledge benchmarks, multiple model families and architectures, model sizes from 0.6B to 16B, data scales from 1B to 5B tokens, multiple baselines, and complementary analyses. KDoS achieves the best average performance and the best result on five of the six benchmarks. The finding that excessive semantic concentration can reduce knowledge-injection efficiency is also potentially useful.

Critical weaknesses

W1: The Knowledge-Density Definition Does Not Reliably Characterize Model Knowledge. The proposed density is computed from the token count and the global mean radius of data embeddings, rather than from the target model's responses or internal state. It therefore cannot indicate which knowledge points are missing or sparse for a particular model, and it is unclear whether density values are comparable across models, embeddings, or domains. Because the target density is selected empirically through downstream sweeps, the definition currently functions more like a task-specific data-selection hyperparameter than an operational measure of model knowledge, limiting its practical utility and generalizability.

W2: The Model's Knowledge Boundary Is Neither Explicitly Modeled nor Directly Evaluated. The paper claims that an optimal density interval consistently maximizes knowledge-boundary expansion. However, it provides neither an explicit model nor an operational measure of the target model's knowledge boundary. Performance differences on factual-QA benchmarks only show that certain density settings yield higher benchmark accuracy; they do not reveal whether previously unknown or sparse knowledge has been expanded, much less whether such expansion has been maximized. The central knowledge-boundary claim is therefore not directly supported by the current evaluation.

Questions for authors

Q1: How is the proposed density expected to reflect model-specific knowledge sparsity, and can the authors demonstrate that it is comparable across target models, domains, and embedding choices?

Q2: The paper claims that an optimal knowledge-density interval consistently maximizes knowledge-boundary expansion. How can the authors demonstrate that it maximizes boundary expansion, given that benchmark performance alone does not directly measure the model's knowledge boundary?

Reviewer_DCGpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how the distribution of synthetic training data affects the effectiveness of knowledge injection into LLMs. The authors argued that existing synthesis methods largely ignore the semantic distribution of generated data, potentially leading to redundant coverage of some knowledge areas and insufficient coverage of others. The paper proposes a framework that introduces a knowledge-density metric to control the concentration of training samples in semantic space.

Critical strengths
  1. The paper points at a real problem. Most synthesis pipelines generate data until they hit a token count and stop, without checking whether some topics are covered over and over while others barely show up. However, there is a limitation in the baseline comparison pointed in the weakness.
  2. Extensive experiments are performed. The authors test models from 0.6B to 16B, three model families (Qwen, Ling, LLaMA), and 1B to 5B tokens of training data.
  3. The idea is easy to understand and use and results show promising improvement over the baselines. It turns “how spread out is my data” down to one number you can set and steer toward, so you don’t need a knowledge graph or a complicated scoring model.
Critical weaknesses
  1. The objective is defined in line 152 as: "The synthetic data are used to inject knowledge into LLM M, and Acc(M, Ssyn) denotes the accuracy of the resulting model on the target test set after training on Ssyn. The goal is to maximize this accuracy ...". Prior work such as ProGen [1] and GOLD [2] directly maximize this accuracy in an end-to-end manner by generating synthetic data and fine-tuning LLM M and providing feedback to maximize the Accuracy of M. GOLD especially undercuts the claim that framing synthesis as distribution optimization is new. It is not clear why a data-only density signal is preferable to student-driven feedback like GOLD and ProGen.

  2. The formal problem in Eq. 2 is never actually solved. Authors write it as an optimization, but in practice they try about seven density values and keep the best one, which is a grid search.

  3. the density seems to be tuned on the test sets. It’s picked by looking at loss on the same benchmarks used to report results, and the baselines don’t get the same tuning. That makes the comparison less fair.

[1] ProGen: Progressive Zero-shot Dataset Generation via In-context Feedback, EMNLP 2022

[2] Generalized Knowledge Distillation via Out-of-Distribution-Guided Language Data Generation, NAACL 2024

Questions for authors

See the weakness section

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.