Think Through a Bottleneck: Hourglass Reasoning for Rigorous Induction
On the map: LLM reasoning → Reasoning benchmarks and evaluation
Abstract (arXiv)
Self-refinement often fails to strengthen few-shot inductive reasoning in large language models. Prompting a model to explicitly state its inferred rule does little on its own. What actually matters is a structurally enforced isolation between reasoning stages, so that information can only pass between them as a compressed symbolic state. We introduce Hourglass reasoning, which enforces strict context isolation between reasoning stages. The frozen LLM acts as a meta-constructor, building for each task a symbolic encoder--decoder: an Induction module compresses the support examples into a schema φ (encoder) and a transient scaffold z; a Deduction module derives rule T (decoder) from these and discards z; an Implementer compiles (φ, T) into artifacts; an error-driven Refiner revises (φ, T) and regenerates artifacts from scratch. Only (φ, T) crosses stage boundaries, so all refinement stays anchored to the rule. We evaluate Hourglass across three benchmarks spanning visual abstraction, hardware synthesis, and textual rule induction, using GPT-5.5 and Gemini 3.1 Pro. On ARC-AGI-2, it raises best-of-5 accuracy by up to 14 points over an iterative-refinement baseline. On ChipBench, it nearly doubles Verilog synthesis accuracy with GPT-5.5, from 31% to 58%. BBEH-Linguini draws on puzzles from the International Linguistics Olympiad, a setting where prior work has shown that explicit verbalization can hurt performance. Hourglass mitigates this tendency, and on Gemini 3.1 Pro, it reverses the effect entirely. Ablations confirm that these gains come from the isolation between stages and the quality of the initial induction, not from prompt wording or the particular symbolic form used. It is how information flows through the reasoning process, rather than the language used to express it, that drives inductive reasoning in frozen LLMs.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 33 of its 56 references that are in the map. First posted on arXiv 13 Jul 2026.
Builds on
33 of 56 references.
Closest ancestors · 5
Same direction, newest first
- 2025UNVEILING: What Makes Linguistics Olympiad Puzzles Tricky for LLMs?15 Aug
- 2025LingBench++: A Linguistically-Informed Benchmark and Reasoning Framework for Multi-Step and Cross-Cultural Inference with LLMs22 Jul
- 2025BIG-Bench Extra Hard26 Feb
- 2025IOLBENCH: Benchmarking LLMs on Linguistic Reasoning8 Jan
- 2023Improving Factuality and Reasoning in Language Models through Multiagent Debate23 May
Other references · 28
- 2022Self-Consistency Improves Chain of Thought Reasoning in Language Modelscited by 5,601
- 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,383
- 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,928
- 2023MetaGPT: Meta Programming for A Multi-Agent Collaborative Frameworkcited by 1,761
- 2018ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustnesscited by 1,430
- 2023Large Language Models Cannot Self-Correct Reasoning Yetcited by 965
- 2016Deep Variational Information Bottleneckcited by 915
- 2023ChatDev: Communicative Agents for Software Developmentcited by 893
- 2023Large Language Models Can Be Easily Distracted by Irrelevant Contextcited by 820
- 2015Deep Learning and the Information Bottleneck Principlecited by 793
- 2025Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMscited by 276
- 2018Hyperspherical Variational Auto-Encoderscited by 188
16 more
- 2025Dynamic Cheatsheet: Test-Time Learning with Adaptive Memorycited by 146
- 2024Large Legal Fictions: Profiling Legal Hallucinations in Large Language Modelscited by 144
- 2023Comparing Humans, GPT-4, and GPT-4V On Abstraction and Reasoning Taskscited by 54
- 2024SimpleToM: Exposing the Gap between Explicit ToM Inference and Implicit ToM Application in LLMscited by 51
- 2021Leveraging Language to Learn Program Abstractions and Search Heuristicscited by 44
- 2024Understanding the Dark Side of LLMs' Intrinsic Self-Correctioncited by 42
- 2019BottleSum: Unsupervised and Self-supervised Sentence Summarization using the Information Bottleneck Principlecited by 39
- 2019Specializing Word Embeddings (for Parsing) by Information Bottleneckcited by 27
- 2025DeepRTL: Bridging Verilog Understanding and Generation with a Unified Representation Modelcited by 23
- 2025Self-Correction Bench: Uncovering and Addressing the Self-Correction Blind Spot in Large Language Modelscited by 19
- 2025Revisiting LLM Reasoning via Information Bottleneckcited by 19
- 2025Product of Experts with LLMs: Boosting Performance on ARC Is a Matter of Perspectivecited by 17
- 2026SkillMaster: Toward Autonomous Skill Mastery in LLM Agentscited by 8
- 2026ChipBench: A Next-Step Benchmark for Evaluating LLM Performance in AI-Aided Chip Designcited by 7
- 2025Corrective In-Context Learning: Evaluating Self-Correction in Large Language Modelscited by 4
- 2025Probing Large Language Models in Reasoning and Translating Complex Linguistic Puzzlescited by 1
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Abstract Activation Spaces for Content-Invariant Reasoning in Large Language Models
- 2025Patterns Over Principles: The Fragility of Inductive Reasoning in LLMs under Noisy Observations
- 2026TIME: Temporally Intelligent Meta-reasoning Engine for Context-Triggered Explicit Reasoning
- 2026Inducing Reasoning Primitives from Agent Traces
- 2023Large Language Models are In-Context Semantic Reasoners rather than Symbolic Reasoners
- 2026A Systematic Study of Small Language Models on Abstract Reasoning Tasks
- 2026In-Context Learning Amplifies a Latent Symbolic Circuit
- 2026Compliance vs. Sensibility: On the Reasoning Controllability in Large Language Models