acceptodds
Under review as a conference paper at ICLR 2027

Mapping Before Generating: IMAP for Inductive Reasoning in Large Language Models

Abstract

Large language models excel at deductive reasoning and context learning, but their inductive reasoning capabilities remain limited by the fact that the models tend to fit instance-level answers rather than abstract reusable general patterns from a finite set of examples. This shortcoming severely hampers the generation of scientific hypotheses, cross-domain generalization, and efficient learning from samples. We propose IMAP, a reinforcement learning-based intelligent mapping framework that integrates the inductive reasoning paradigm into model inference. We designed a paradigm mapping structure comprising four core elements: chains of thought (CoT), cases, patterns, and rationality. We also propose the RL Paradigm Model (RLP) for paradigm generation. By using inductive reasoning as an input cue, we guided multiple large models to generate an average of 270 results. Comparative experiments demonstrate that combining input cues with inductive reasoning performs well across most models and enhances the quality of generated results. By framing inductive reasoning as a learnable paradigm-mapping process, our work unlocks its generative potential and offers a new perspective for understanding and deconstructing the reasoning mechanisms of large language models.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Inductive thinking paradigm, large-scale language modeling, BBH benchmark assessment, jina-score, reliability assessment

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ECnKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The author claims that current large language models perform poorly at inductive reasoning. The author proposes IMPA, an RL-based mapping framework that incorporates CoT, cases, patterns, and rationality. The author also proposes RLP for paradigm generation. The author conducts experiments and claims that their paradigm improves the quality of the generated results. The paradigm first extracts cases and patterns from the initial CoT to form a graph. Then, a verification module provides a reward based on soft consistency.

Critical strengths
  1. The reward module can be self-verified without relying on an external reward model.
Critical weaknesses
  1. I think the author, in general, should be more careful about writing. e. g., in the first paragraph of the introduction, the author makes some very strong claims about the limitations of LLMs (tend to be suited for instance-level answers rather than autonomously inducting abstract, reusable, and general patterns) or the underlying mechanism of LLMs (treat CoT as the overall trajectory of policy optimization, without explicit knowledge abstraction). The authors should either include at least some toy experiments to support the claim or soften the tone.
  2. The author chooses BigBenchHard and other general reasoning benchmarks. Also, in the related work, the author seems to care only about general knowledge-based reasoning rather than inductive reasoning. The author should consider either properly scoping the paper as inductive reasoning (consider changing the benchmark to inductive reasoning-specific ones such as InductiveBench[1]) or broadening the scope of the paper to be a general knowledge-reasoning paper.

[1] InductionBench: LLMs Fail in the Simplest Complexity Class

Questions for authors

How is the IAMP framework working at inference time? Does it still have the outer-loop section, and is there any difference between the training-time behavior?

Reviewer_Z4nLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes IMAP, a reinforcement learning framework for inductive reasoning in Large Language Models. The method runs a loop to help LLMs extract reusable rules from examples, instead of simply fitting instance-level answers. The policy first writes a CoT, then an LLM operator splits the reasoning into cases, which will be further abstracted into patterns by another operator. The policy is trained with PPO under structurally informed rewards. Experiments use Llama-3.2-1B on BBH and Llama-3.1-8B on MATH, GSM8K, HHS, LegalBench and PubMedQA. The paper reports 37.2% on BBH and gains over RAFT-RL and ToT, and concludes that framing induction as pattern mapping could improve accuracy and interpretability.

Critical strengths
  1. The idea of keeping a memory of cases and patterns, and using it in both the prompt and the reward, is a reasonable design for a process reward, while most RL baselines only have accuracy reward for the final answer.
  2. The paper provides a bootstrap test and a code link for reproducing the work.
Critical weaknesses
  1. It is confusing that the same number of 37.21% is reported for two methods. Table 1 gives 37.21% for RL-trained IMAP on Llama-3.2-1B. Table 3 and Figure 4(c) give the same 37.21% for “Inductive (Ours)”, which is described as a simple prompting method with no training.
  2. The main gain is not significant, for example, the gain over AbstRaL is ~ 0.1%, the gain over RAFT-RL is ~1%, at a higher inference cost (19.3 vs 12.4 LLM calls per task).
  3. All RL experiments use only Llama-3.2-1B (BBH) or Llama-3.1-8B (other benchmarks), one model per benchmark. No other model family is tested, and no backbone is run on all benchmarks, so it is unclear to judge if the method transfers on different models.
Questions for authors
  1. Why is 37.21% reported both for RL-trained IMAP (Table 1) and for the prompting-only “Inductive” method (Table 3, Figure 4c)? Which one is correct?
  2. Which model generates the cases and patterns during training, and in what language? The released dataset is in Chinese and its answer column is empty.
Reviewer_PdM1OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes IMAP, a reinforcement-learning framework that periodically abstracts LLM reasoning traces into structured cases and textual patterns, retrieves these patterns to guide subsequent reasoning, and rewards consistency with them; while it reports improvements on several reasoning benchmarks, the empirical evidence does not clearly establish the paper’s stronger claim that the framework meaningfully learns general inductive reasoning abilities.

Critical strengths
  • The paper targets an important problem on moving beyond instance-level reasoning toward inductive reasoning.
  • The paper includes useful ablations over major components.
  • IMAP provides interesting reasoning structure.
Critical weaknesses
  • The paper makes a broad claim about improving inductive reasoning, but the experiments mostly measure performance on standard reasoning benchmarks. Hence it's unclear whether IMAP genuinely improves inductive reasoning.
  • The empirical advantage of IMAP is not consistent across benchmarks and comparisons. Improvements over the strongest baselines are small in several settings, and the per-task results in Table 2 are mixed. Overall, the results do not clearly demonstrate a consistent advantage of the proposed method over existing approaches.
  • The full RL-trained IMAP method is evaluated only on two Llama model variants. While Table 3 evaluates models from additional families such as Yi and Qwen, these experiments only test the proposed inductive prompting paradigm rather than the full RL trained IMAP pipeline. The experiments provide limited evidence that the proposed training method generalizes across model families.
  • The experimental setup does not sufficiently reflect the current state of LLM reasoning research. Most of the primary baselines are older prompting or search-based methods, such as CoT, Self-Consistency, ToT, and RAP, while the paper does not compare against more recent prominent reasoning-model training methods. Moreover, evaluation focuses heavily on benchmarks such as GSM8K, MATH, and BBH, which are increasingly saturated by modern reasoning models and may not adequately demonstrate advantages under current more challenging reasoning settings.
Questions for authors

How does IMAP relate to recent advances in RL-based LLM reasoning, and what advantage does explicit Case/Pattern induction provide beyond reasoning capabilities that can emerge directly from modern reasoning-model training?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.