PaperlayerView exact revision on arXiv

Source-linked AI summary

LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure

Fanfei Li, Jana Zeller, Manuel Prada-Corral, Thaddäus Wiedemer, Prasanna Mayilvahanan, Ryan Cotterell, Wieland Brendel

arXiv:2608.13545v1cs.CLcs.AIcs.LG

TL;DR

Unknown prior exposure in web-scale training makes it difficult to determine whether language models genuinely acquire capabilities. The paper creates a K–5-controlled corpus and model, finding that post-training and in-context learning improve in-scope use but do not substantially extend out-of-scope performance.

  • Problem

    Unknown prior exposure in heterogeneous web-scale corpora makes it difficult to determine whether later methods elicit existing knowledge or produce genuine capability growth.

  • Method

    The paper constructs an 88B-token K–5 corpus and trains LITTLELEARNER, a 5B model, to create a controlled sandbox for studying restricted knowledge exposure.

  • Results

    Increased model size, post-training, and in-context learning do not improve out-of-scope performance far beyond training exposure, but can improve performance within scope and at the boundary.

  • Takeaways & Limitations

    LITTLECURRICULUM and LITTLELEARNER provide a controlled sandbox for studying knowledge acquisition, representation, and utilization within a defined training scope.

  • Takeaways & Limitations

    The developmental framing controls exposure but does not model human-like development, because human and machine learners acquire language, concepts, and procedures through different mechanisms.

Abstract

from arXiv · show

Modern language models are trained on heterogeneous web-scale text corpora. Consequently, studying knowledge and skill acquisition is difficult, as prior exposure to related content is hard to characterize. To address this challenge, we introduce LITTLECURRICULUM, a curated 88B-token pretraining corpus tailored to U.S. elementary school material, explicitly excluding concepts, facts, and vocabulary taught above Grade 5. Training a 5B-parameter LLM from scratch on LITTLECURRICULUM yields LITTLELEARNER, a model with sufficient language competence for open-ended evaluation, yet with clear knowledge and capability boundaries mapped to interpretable curriculum guidelines. We release LITTLECURRICULUM and LITTLELEARNER as a developmentally restricted sandbox to study how models acquire, represent, and use data under a well-defined training scope. We illustrate the sandbox's utility in a first suite of experiments on injecting new knowledge through post-training and in-context learning. These methods let LITTLELEARNER better utilize existing knowledge, but do not raise out-of-scope capabilities. Our findings underscore the value of this controlled environment for future investigations.

1 Introduction

The paper addresses the difficulty of characterizing language models’ prior knowledge by constraining pretraining exposure to a pedagogically defined K–5 curriculum. It introduces LITTLECURRICULUM and LITTLELEARNER as a controlled sandbox for studying knowledge acquisition within this scope.

  • Motivation: Web-scale pretraining makes prior knowledge difficult to characterize, while contamination and brittle generalization can distort observed model behavior.Targeted data patches may also be needed to repair missing skills.
  • Motivation: The authors complement benchmark-based novelty evaluation by constraining the training distribution rather than redefining novelty for each evaluation.They use developmental boundaries because age-level curricula provide a pedagogically grounded proxy for exposed concepts, vocabulary, and skills.
  • Contributions: 88B tokens: LITTLECURRICULUM is a curated dataset containing only U.S. elementary school K–5 content.It is constructed with a scalable filtering pipeline.
  • Contributions: 5B parameters: LITTLELEARNER is trained from scratch on LITTLECURRICULUM to form, with the dataset, a sandbox for probing data acquisition, representation, and utilization.The sandbox has a well-defined training scope.
  • Findings: Increased model size, post-training, and in-context learning improve within-scope and boundary performance but do not substantially improve out-of-scope performance beyond training exposure.These are the paper’s first experiments with LITTLELEARNER.

2 Related Work

Related work studies knowledge boundaries after training, emphasizes the dominance of pretraining data over downstream behavior, and reduces distributional confounds through evaluation benchmarks or constrained corpora. LittleLearner instead specifies the knowledge boundary in advance by controlling the training distribution.

  • Knowledge Boundaries in LLMs: Existing methods infer an already-trained model’s knowledge boundary through self-knowledge, unanswerable-question detection, confidence calibration, retrieval-augmented behavior shifts, and prompt sensitivity.These methods operate over an opaque pretraining distribution and characterize what a fixed LLM does or does not know after training.
  • Pretraining data dominates downstream behavior: Pretraining distribution is the primary determinant of generalization and downstream capability, while architecture and optimization play comparatively minor roles.Analyses of reinforcement learning further suggest that observed gains often elicit reasoning patterns already present in pretraining rather than introduce genuinely new capabilities.
  • Evaluation beyond training distribution: Evaluation benchmarks such as Humanity’s Last Exam, GPQA, MMLU-Pro, and MATH-B minimize overlap with web-scale pretraining but provide only indirect control.Because the underlying training distribution remains opaque, these benchmarks must be continuously refreshed to stay informative.
  • Constrained pretraining distributions: BabyLM restricts corpus quantity rather than conceptual scope, leaving advanced material accessible, while concurrent work imposes a temporal cutoff without constraining conceptual complexity.These approaches directly constrain pretraining distributions but differ from LittleLearner’s advance control of conceptual scope.

3 Constructing the LITTLELEARNER-LITTLECURRICULUM Sandbox

The LITTLELEARNER-LITTLECURRICULUM sandbox is built by precision-first filtering of FineWeb-Edu to enforce a K–5 exposure boundary, then training and evaluating a 5B-parameter model on language, mathematical, and factual tasks. Validation shows strong in-scope performance and increasing unfamiliarity or degradation beyond the K–5 scope.

  • Pipeline validation: 0% of Beyond-K–5 content was retained on CommonCoreText while approximately 35% of K–5 content was preserved, reflecting a precision-over-recall design.On WeeBit, the filter retained 2.48% of 6,000 Beyond-K–5 passages, with manual inspection identifying genuinely out-of-scope concepts in only three passages, or 0.05% of that split.
  • Corpus construction: LITTLECURRICULUM is a K–5-targeted FineWeb-Edu subset constructed with multi-stage filters over knowledge, syntactic complexity, reasoning, and disproportionately out-of-scope terms.The pipeline combines AoA-based age alignment, FastText and ModernBERT classifiers, symbolic notation filtering, and final term-based filtering.
  • Model training: A 5B-parameter LITTLELEARNER model was trained from scratch for 100 hours on 8 NVIDIA B200 GPUs using the Qwen3 architecture.The model was trained on the constructed datasets and evaluated against UNFILTERED and Gemma 2B reference models.
  • Capability validation: LITTLELEARNER’s bits-per-byte rises with increasing CLEAR passage hardness, while UNFILTERED and Gemma 2B remain comparatively flat.The divergence is attributed to pretraining data because UNFILTERED shares LITTLELEARNER’s training recipe.
  • Capability validation: In Beyond-K–5 mathematical content, LITTLELEARNER exhibits growing unfamiliarity while unrestricted reference models remain stable; in K–5, its BPB is similar to the reference models.The comparison uses student turns from CoMTA to test mathematical familiarity.
  • Capability boundaries: LITTLELEARNER performs strongly on in-scope language, math, and factual tasks but shows a sudden factual-knowledge drop and disproportionately large higher-grade mathematical-reasoning decline beyond K–5.Increasing the sampling budget does not eliminate the higher-grade mathematical performance gap.

4 Probing Capabilities with LITTLELEARNER

Across model scaling, post-training, and in-context learning, LITTLELEARNER gains capabilities within its K–5 exposure but shows limited improvement on Beyond-K–5 tasks. In the tested settings, the pretraining filter sets the effective capability ceiling.

  • Model scaling: Scaling improves reasoning within K–5 exposure but does not meaningfully extend LITTLELEARNER’s capability boundary.Larger models show little improvement on problems requiring advanced capabilities outside the exposure.
  • Post-training: Post-training significantly boosts in-scope K–5 capabilities but fails to recover Beyond-K–5 performance, even with out-of-scope data.The study compares SFT followed by GRPO for LITTLELEARNER with a compute-matched unfiltered baseline.
  • In-context learning: Few-shot in-context learning slightly improves K–5 performance but does not unlock new Beyond-K–5 reasoning capabilities.Adding explanations has no effect on performance in the tested setup.
  • Overall findings: All three interventions amplify capabilities within K–5 exposure while providing limited Beyond-K–5 gains, leaving the pretraining filter as the effective capability ceiling.This pattern holds across scaling, SFT+GRPO post-training, and in-context learning in the tested settings.

5 Exploration Avenues: Beyond the K–5 Boundary

The LITTLECURRICULUM–LITTLELEARNER sandbox enables controlled study of capability extension, continual learning, interpretability, calibration, and memory-based mechanisms under explicit K–5 exposure constraints. Its developmental framing supports educational and human–model comparisons only as exposure control, not as a model of human learning.

  • Section rationale: The authors frame the LITTLECURRICULUM–LITTLELEARNER sandbox as a playground for studying how language models acquire and extend knowledge under explicit competency constraints.The controlled exposure boundary is the basis for the exploration avenues described in this section.
  • RL, discovery, and extrapolation: The sandbox separates capabilities emerging under reinforcement learning from latent pretraining knowledge because LITTLELEARNER’s prior is restricted to K–5 material.This enables studying capabilities such as multi-digit arithmetic or multi-step algebraic reasoning as they emerge under RL.
  • Continual learning, interpretability, and calibration: Explicitly specified exposure makes behavioral and representational changes more directly relatable to concepts introduced during continual learning, interpretability, and calibration studies.Web-scale pretraining otherwise blurs generalization, memorization, and latent knowledge.
  • Educational science and human–model comparison: LITTLELEARNER’s developmental boundary controls exposure but does not model human children, whose learning mechanisms differ from machine learners.The model can sometimes perform better in downstream skills than in prerequisites.
  • Interactive and memory-based learning: The sandbox’s known prior exposure enables direct evaluation of retrieval, external memory, self-reflection, exploration, and multi-agent collaboration.These directions test integrating out-of-domain content, recombining latent capabilities without new data, and tracking newly introduced information.

6 Conclusion

LITTLECURRICULUM and LITTLELEARNER provide a developmentally constrained, controlled sandbox for studying knowledge acquisition. Post-training and in-context learning do not unlock reasoning patterns or information from out-of-scope domains.

  • 6 Conclusion: LITTLECURRICULUM is an 88B-token developmentally constrained corpus, paired with LITTLELEARNER, a 5B model for controlled knowledge-acquisition studies.The model is small enough for academic feasibility while supporting coherent production.
  • 6 Conclusion: Post-training and in-context learning fail to unlock reasoning patterns or information from out-of-scope domains.These knowledge-injection methods do not overcome the model’s developmental constraints.

A Further Details on the Dataset Filtering Pipeline · A.1 Rule Based Pre-Filtering

The dataset filtering pipeline evaluates rule-based metrics for sorting FineWeb-Edu and identifies Age-of-Acquisition (AoA) as the clearest signal for grade-band alignment. Word-frequency metrics are then assessed for filling AoA coverage gaps and reducing out-of-vocabulary words.

  • A.1 Rule Based Pre-Filtering: AoA gives the highest signal among the tested rule-based approaches for pre-sorting FineWeb-Edu.The comparison includes readability, linguistic, and frequency metrics.
  • A.1 Rule Based Pre-Filtering: The evaluation tests Flesch-Kincaid and Dale-Chall readability metrics alongside Type Token Ratio and average sentence and word length.
  • A.1 Rule Based Pre-Filtering: The frequency-based candidates include New Academic Word List scores and Wikipedia-derived IDF scores.
  • A.1 Rule Based Pre-Filtering: 2.5% of all words remain OOV when SUBTLEX-US frequencies are used, despite their strongest alignment with AoA scores.The frequency metric is evaluated as a way to reduce OOV words from AoA alone.
  • A.1 Rule Based Pre-Filtering: The AoA-plus-Zipf analysis reports the most frequent words absent from both Zipf and AoA in a subset of FineWeb-Edu.
  • A.1 Rule Based Pre-Filtering: Figure 10 identifies Age-of-Acquisition as the best fit for underlying grade-band classification, spanning K–5, middle school, high school, and out-of-scope categories.K5 denotes K–5, K8 middle school, K12 high school, and OOS material beyond school level.
  • A.1 Rule Based Pre-Filtering: At higher word coverage, Word Freq Zipf provides the best trade-off between increased word coverage and alignment to AoA.

A.2 LLM-J … A.5 CommonCoreText

The filtering pipeline combines LLM-J grade-band classification, symbolic cleanup, and frequency-based sampling to enforce the K–5 exposure boundary. Evaluation on held-out CommonCoreText shows that filtering preferentially removes longer and more lexically complex passages while retaining simpler in-scope content.

  • A.2 LLM-J: LLM-J classifies snippets into four grade bands using majority voting across optimized prompts targeting language, context, and subject complexity.The subject prompt aggregates math, life science, physical science, and earth/space science signals; prompts are optimized independently with OpenEvolve.
  • A.2 LLM-J: Single-judge Gemini labeling achieves the highest accuracy and lowest under-classification, while safety adjustment reduces under-classification but harms reliability and recall.Confidence-weighted voting does not improve over uniform majority voting, and in-context learning matches the no-ICL setup at roughly 2× the cost.
  • A.3 Symbolic Filtering: ~0.1% is the approximate symbolic-pattern removal rate after ModernBERT classification, with polynomial expressions, exponent notation, and multi-variable equations removed most often.The rule-based filter targets equations, exponents, roots, inequalities, functions, and calculus operators because formal notation is sparse in the corpus.
  • A.3 Symbolic Filtering: The symbolic stage functions as a narrow cleanup pass rather than a general mathematics detector because formal notation, not advanced vocabulary, drives its removals.The largest removed categories are cases where Beyond-K–5 content relies on formal notation.
  • A.4 Frequency Sampling: The final sampling stage tightens the exposure boundary by ranking terms disproportionately associated with Beyond-K–5 material and downsampling candidate documents containing high-scoring terms.Because the retained pool contains roughly 100B tokens versus an 80B-token training target, frequency statistics guide preferential exclusion rather than uniform subsampling.
  • A.5 CommonCoreText: 460 held-out CommonCoreText passages span grade levels, science and literature domains, and instructional and canonical narrative materials for evaluating retention and specificity.The benchmark is sampled from textbooks and reading materials aligned with CCSS and remains excluded from classifier training.
  • A.5 CommonCoreText: 115 to 205 words is the median-length change from retained to removed K–5 passages, which are consistently longer and more lexically complex.This pattern matches the pipeline’s precision-oriented design, favoring simpler and clearly in-scope content.

A.6 Independent Leakage Checks … Step 2: Frequency Sampling

The paper checks leakage independently and illustrates how successive filtering stages remove out-of-scope material. External-corpus inspection finds limited leakage, while examples show filters targeting age, topic, sentence complexity, symbols, and frequency.

  • A.6 Independent Leakage Checks: 0.05% of the full Beyond-K–5 split consists of three WeeBit passages containing clearly out-of-scope concepts: fractions, exterior angles, and negative temperatures.The retained samples also included web-scraping artifacts and simple narrative prose.
  • A.6 Independent Leakage Checks: The WeeBit evaluation and corpus-level terminology audit provide complementary evidence that Beyond-K–5 leakage through the filtering pipeline is limited.The audit compiles 126 Grade-6+ technical n-grams from mathematics, English Language Arts, and science standards.
  • A.7 Examples of Dropped Samples per Filtering Stage: AoA pre-filtering removes documents clearly outside the knowledge cutoff, including texts about neuromodulation, neurology, cardiology, and cardiac electrophysiology.The examples include detailed medical descriptions and emergency-room cardiology discussion.
  • Step 2: FastText Classification: FastText classification discards longer-sentence text and documents discussing out-of-scope material such as brewing equipment and the Gardner algorithm.The Gardner example concerns identifying a counterfeit coin with a balance scale.
  • Step 3: ModernBERT Classification: ModernBERT classification removes documents containing medical guidance or topics, including nasal foreign bodies and benign prostatic hyperplasia.The examples include emergency advice and treatment options involving lifestyle changes, medicines, and surgery.
  • Step 2: Frequency Sampling: Frequency sampling examples include documents about humanism, Darwin Day, and speculative Atlantean history.The examples show sampled material spanning science-and-reason advocacy and historical or prehistorical claims.

B Details on Model Training … C.2.1 Filtering MathCAMPS Standards and Questions

The paper trains LITTLELEARNER from scratch on K–5-filtered data with controls against tokenizer leakage and selects the best validation checkpoint. Its MathCAMPS evaluation filters unreliable items and shows that K–5 and Beyond-K–5 questions have closely matched readability metrics.

  • B Details on Model Training: LITTLELEARNER uses the Qwen3-dense architecture and is trained from scratch with Megatron-Core, 8-way layer-wise sharded Muon, BF16 storage, and MXFP8 computation.The cooloff mixture contains 91% K–5-filtered pretraining data, 5% K–5-rewritten mathematical reasoning data, and 2% K–5-rewritten mathematical instruction-tuning data.
  • B Details on Model Training: A custom tokenizer trained only on LITTLECURRICULUM prevents Beyond-K–5 leakage through tokenization, while single-digit number tokenization raises pass@1 but leaves LITTLELEARNER near floor.In the matched 2B ablation, Beyond-K–5 MathCAMPS pass@1 increases from 1.1% to 1.8% for LITTLELEARNER and from 4.6% to 10.3% for UNFILTERED.
  • B Details on Model Training: The model retains the checkpoint with the lowest validation bits-per-byte after monitoring a held-out validation stream throughout pretraining.This checkpoint is used for downstream post-training and evaluation.
  • C.1 Mathematical Reasoning with CoMTA: CoMTA is manually labeled Elementary when its dialogs correspond to K–5 material, while topics such as matrix operations, negative exponents, trigonometry, and calculus are treated as Beyond-K–5.The supplied examples include decimal multiplication, fraction addition, and counting to ten as K–5-aligned topics.
  • C.2.1 Filtering MathCAMPS Standards and Questions: MathCAMPS filtering removes standards with fewer than 30 unique gold answers because low-entropy answer distributions can inflate pass@k independently of reasoning ability.The removed standards are predominantly multiple-choice items or questions involving perfect squares, cubes, and cube roots.
  • C.2.1 Filtering MathCAMPS Standards and Questions: Questions are discarded when the gold answer appears verbatim in the prompt, because models can restate the prompt and receive credit from the regex-based grader without solving.Performance on these questions is up to 23× higher than on otherwise comparable questions within the same standard.
  • C.2.1 Filtering MathCAMPS Standards and Questions: The evaluation excludes CCSS standard 6.EE.B.7 because its MathCAMPS problems resemble grade-3 addition-and-subtraction word problems despite being labeled grade 6.The paper also validates comparable readability between K–5 and Beyond-K–5 partitions to distinguish conceptual from purely linguistic shifts.
  • C.2.1 Filtering MathCAMPS Standards and Questions: Readability is similar across partitions: Flesch–Kincaid grade is 5.28 for K–5 versus 5.37 for Beyond-K–5, and mean sentence length is 10.7 versus 10.1 words.Mean word length is 4.29 versus 4.20 characters, while mean question length is 28 versus 23 words.

C.2.2 Performance Across Learning Paths · C.2.3 Pass@k Performance on MathCAMPS for LITTLELEARNER and UNFILTERED

LITTLELEARNER follows UNFILTERED closely within K–5 but diverges on higher-grade arithmetic and Beyond-K–5 evaluation, with performance along prerequisite chains remaining non-monotonic. Increasing sampling budgets does not recover the gap, as both models’ pass@k gains flatten well before k=1024.

  • C.2.2 Performance Across Learning Paths: Across Grades 3–6, LITTLELEARNER and UNFILTERED track closely, but they diverge in a framing-dependent way at Grade 7.Figure 15 evaluates both models at pass@1 along CCSS division and division/multiplication learning paths.
  • C.2.2 Performance Across Learning Paths: More K–5-aligned framing improves LITTLELEARNER’s performance but does not close its gap to UNFILTERED.
  • C.2.2 Performance Across Learning Paths: Performance along CCSS prerequisite chains is non-monotonic: LITTLELEARNER is fluent on 3.OA.C.7, drops on 4.NBT.B.6, and recovers on 5.NBT.B.6.Multi-digit multiplication is similarly jagged: 4.NBT.B.5 is easier than 5.NBT.B.5.
  • C.2.2 Performance Across Learning Paths: Scaling model parameters improves performance within K–5 and at the boundary, but does not recover Beyond-K–5 performance well.
  • C.2.3 Pass@k Performance on MathCAMPS for LITTLELEARNER and UNFILTERED: At pass@1024, LITTLELEARNER retains essentially all of UNFILTERED’s competence in Grades K–3, a comparable share through Grades 4–6, and about two-thirds at Grade 8.The Grade 8 comparison has no overlapping confidence intervals, while performance degrades more sharply in Beyond-K–5.
  • C.2.3 Pass@k Performance on MathCAMPS for LITTLELEARNER and UNFILTERED: Both models’ pass@k curves flatten well before k=1024, with gains slowing noticeably by k≈100 and becoming essentially zero thereafter in every panel.The evaluation uses direct prompting under high sampling budgets to increase per-problem output diversity.

C.3 Details on Scaling Experiments … C.5.3 Comparing Different Few-Shot Examples

The experiments examine scaling at the K–5 boundary, post-training data and procedures, and in-context learning formats. Results indicate fuzzy boundary effects near K–5, while natural-prose demonstrations are the only tested format that improves over zero-shot performance.

  • C.3 Details on Scaling Experiments: Scaling shows some effects in grade bands near the fuzzy K–5 versus Beyond-K–5 boundary, but no effect at grade 8.The grade-8 result is well within Beyond-K–5.
  • C.4 Details on Post-Training Experiments: SFT uses 108k grade-school math problems and 526k reading-comprehension pairs, totaling ∼46M tokens for one epoch.A small fraction of the chat data is also folded into pretraining’s final learning-rate decay phase.
  • C.4 Details on Post-Training Experiments: GRPO trains on grade-stratified synthetic problems and filtered GSM8K, re-banding the pool every 150–300 steps to problems the current policy solves 1–15 times out of 16.The synthetic problems are checked for zero overlap with the evaluation set.
  • C.5.1 Question and Explanation Construction: Questions are generated at three per CCSS capability with numerically verified answers, while explanations provide grade-appropriate procedures using only grade-5-or-below operations and vocabulary.Explanations are anchored to the standard, grade, and assumed K–5 prerequisites.
  • C.5.1 Question and Explanation Construction: The generated explanations avoid generic variable placeholders and follow grade-5-style numbered steps with plain ASCII arithmetic and a/b fraction notation.Case-based procedures are instructed to state the relevant cases explicitly.
  • C.5.2 Creating Hand-Authored Natural-Prose CoT: For 43 K–5 CCSS anchor standards, the hand-authored natural-prose demonstrations use an orienting sentence, connective computation steps, and a parseable final Answer: line.The solutions were authored rather than reused from model-generated chain-of-thoughts.
  • C.5.3 Comparing Different Few-Shot Examples: The in-context evaluation compares compact algebraic, natural-prose, and Q/A-only demonstrations while keeping the question sets identical.The compact algebraic and Q/A-only formats have mean lengths of 99 and 41 characters, while natural prose averages 374 characters.
  • C.5.3 Comparing Different Few-Shot Examples: Only natural-prose shots improve over the zero-shot direct baseline; every other format strictly hurts, and adjusting output length does not improve performance.None of the tested in-context-learning shots recover Beyond-K–5 performance.

D More Qualitative Examples · E Existing Assets, Licenses, and Terms of Use

Qualitative comparisons show that LITTLELEARNER explains in-scope concepts accurately but produces coherent, systematically incorrect explanations for unfamiliar concepts. The paper also documents existing assets, their uses, licensing terms, and release inclusion status.

  • D More Qualitative Examples: LITTLELEARNER and UNFILTERED are compared on prompts spanning increasing conceptual difficulty, including K–5 and Beyond-K–5 concepts.These comparisons are presented as qualitative base-model examples in Table 6.
  • D More Qualitative Examples: On in-scope prompts, LITTLELEARNER gives accurate explanations with simplified, progressively structured reasoning characteristic of a spiral curriculum.Its responses are often less detailed than the general model’s while remaining correct and developmentally appropriate.
  • D More Qualitative Examples: Compared with the general model, LITTLELEARNER is often less detailed on in-scope prompts but remains correct and developmentally appropriate.The responses reflect concepts introduced early and refined over time.
  • D More Qualitative Examples: On out-of-scope prompts, LITTLELEARNER produces structured yet incorrect explanations rather than expressing uncertainty.The outputs systematically project unfamiliar concepts onto familiar reasoning patterns learned during training.
  • D More Qualitative Examples: For example, LITTLELEARNER reinterprets Schrödinger’s cat as a literal cat with fabricated attributes.This illustrates that the model does not retrieve correct knowledge for unfamiliar concepts.
  • D More Qualitative Examples: The resulting outputs are coherent but incorrect because LITTLELEARNER lacks the underlying concepts and explains them using available reasoning primitives.The model nevertheless attempts to construct explanations from what it learned during training.
  • E Existing Assets, Licenses, and Terms of Use: Table 7 summarizes existing assets used in the work, their roles, license terms, and whether they are included in the release.Version information, access dates, and source URLs are provided through corresponding citations or release documentation.
  • E Existing Assets, Licenses, and Terms of Use: Unless otherwise indicated, third-party datasets, models, codebases, and curriculum standards are not redistributed in the released artifacts.“Included” covers the asset itself or a derived version of it.
Loading 2608.13545v1…