Pith. sign in

REVIEW 15 cited by

On Memorization of Large Language Models in Logical Reasoning

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2410.23123 v2 pith:W3AK36ZL submitted 2024-10-30 cs.CL

classification cs.CL
keywords reasoningmemorizationllmspuzzlesfine-tuninglogicalperformancebenchmarks
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Large language models (LLMs) achieve good performance on challenging reasoning benchmarks, yet could also make basic reasoning mistakes. This contrasting behavior is puzzling when it comes to understanding the mechanisms behind LLMs' reasoning capabilities. One hypothesis is that the increasingly high and nearly saturated performance on common reasoning benchmarks could be due to the memorization of similar problems. In this paper, we systematically investigate this hypothesis with a quantitative measurement of memorization in reasoning tasks, using a dynamically generated logical reasoning benchmark based on Knights and Knaves (K&K) puzzles. We find that LLMs could interpolate and memorize the training puzzles (achieving near-perfect accuracy) after fine-tuning, yet they struggle with slight variations of these puzzles. On the other hand, we show that while fine-tuning leads to heavy memorization, it also consistently improves generalization performance. Through in-depth analyses with perturbation tests, cross difficulty-level transferability, probing model internals, and fine-tuning with wrong answers, we establish that LLMs develop reasoning skills on K&K puzzles alongside memorization. Finally, our analysis based on a per-sample memorization score sheds light on how LLMs switch between reasoning and memorization when solving logical puzzles. Our code and data are available at https://memkklogic.github.io.

Discussion (0). Sign in to comment.

Forward citations

Cited by 15 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond

    cs.AI 2025-05 conditional novelty 7.0 of 10

    A synthetic logic dataset with verifiable answers, trained via reinforcement learning, improves LLM logical reasoning and generalizes to math and coding.

  2. Enough is as good as a feast: A Comprehensive Analysis of How Reinforcement Learning Mitigates Task Conflicts in LLMs

    cs.CL 2026-07 conditional novelty 6.0 of 10

    RL-trained LLMs keep most of their skills after weight merging, while SFT-trained LLMs drop about 19% on average, because RL keeps parameter updates smaller and more task-compatible.

  3. RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments

    cs.CL 2025-11 conditional novelty 6.0 of 10

    Training a 1.5B language model on 400 adaptively-difficult, procedurally generated, verifiable environments improved average reasoning-benchmark performance by 3.37% absolute, compared with 0.49% from continuing origi...

  4. Understanding the Ability of LLMs to Handle Character-Level Perturbation

    cs.CL 2025-10 conditional novelty 6.0 of 10

    LLMs remain surprisingly accurate on math and coding when invisible Unicode noise is inserted after every character, with robustness driven by implicit internal denoising and, for some models, explicit rewriting in ch...

  5. Throttling Web Agents Using Reasoning Gates

    cs.AI 2025-09 conditional novelty 6.0 of 10

    Rebus-based reasoning gates, puzzles built from random word/domain clue sets, impose token costs on LM web agents that are up to 9.2x the generator's cost.

  6. Can One Domain Help Others? A Data-Centric Study on Multi-Domain Reasoning via Reinforcement Learning

    cs.AI 2025-07 conditional novelty 6.0 of 10

    Training a Qwen2.5-7B model with GRPO on math and puzzle data improves both domains, code transfer depends on the starting model, and template or reward mismatches sharply hurt performance.

  7. ABench-Physics: Benchmarking Physical Reasoning in LLMs via High-Difficulty and Dynamic Physics Problems

    cs.LG 2025-07 conditional novelty 6.0 of 10

    A new physics benchmark with static and dynamically varied numeric problems shows top LLMs solve at most 43 percent of the static set and drop sharply when problem constants change.

  8. OIBench: Benchmarking Strong Reasoning Models with Olympiad in Informatics

    cs.AI 2025-06 conditional novelty 6.0 of 10

    A private, contamination-resistant benchmark of 250 olympiad-level programming problems shows top reasoning models reaching about 36% solve rates, far above conventional models.

  9. SV-TrustEval-C: Evaluating Structure and Semantic Reasoning in Large Language Models for Source Code Vulnerability Analysis

    cs.SE 2025-05 conditional novelty 6.0 of 10

    Large language models perform poorly on a new C-code vulnerability benchmark, indicating they rely on pattern matching rather than genuine reasoning.

  10. Towards Revealing the Effectiveness of Small-Scale Fine-tuning in R1-style Reinforcement Learning

    cs.LG 2025-05 conditional novelty 6.0 of 10

    Re-distilling a model's own RL-trained policy into 1K SFT samples reproduces RL accuracy at a fraction of the compute.

  11. Adaptive Multi-Agent Reasoning via Automated Workflow Generation

    cs.AI 2025-07 conditional novelty 5.0 of 10

    Automated workflow generation and iterative prompt refinement let a standard GPT-4.1 model outperform state-of-the-art reasoning models on a revised riddle benchmark.

  12. Answer-Centric or Reasoning-Driven? Uncovering the Latent Memory Anchor in LLMs

    cs.CL 2025-06 conditional novelty 5.0 of 10

    Across ten LLMs, masking the final answer inside a complete reasoning chain causes a 26.9-point accuracy drop, evidence that models anchor to answers, not reasoning templates.

  13. The Avengers: A Simple Recipe for Uniting Smaller Language Models to Challenge Proprietary Giants

    cs.CL 2025-05 reject novelty 5.0 of 10

    Clustering-based routing plus self-consistency voting among ten 7B open models reportedly outranks GPT-4.1 and GPT-4.5 on average over 15 diverse benchmarks.

  14. Infinite Sampling: Efficient and Stable Grouped RL Training for Large Language Models

    cs.LG 2025-06 conditional novelty 4.0 of 10

    A GRPO decoding framework that cuts memory via micro-batched KV-cache reuse and improves decoding-round efficiency with predicted-length scheduling, at the cost of serialization.

  15. Beyond Frequency: The Role of Redundancy in Large Language Model Memorization

    cs.LG 2025-06 reject novelty 4.0 of 10

    LLM memorization is concentrated in low-redundancy (high-entropy) samples, which are roughly twice as fragile under prefix perturbation as high-redundancy samples, according to the paper's measurements on Pythia models.

Pith tools