REVIEW 44 cited by
To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic 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
Signed reviews
read the original abstract
Chain-of-thought (CoT) via prompting is the de facto method for eliciting reasoning capabilities from large language models (LLMs). But for what kinds of tasks is this extra ``thinking'' really helpful? To analyze this, we conducted a quantitative meta-analysis covering over 100 papers using CoT and ran our own evaluations of 20 datasets across 14 models. Our results show that CoT gives strong performance benefits primarily on tasks involving math or logic, with much smaller gains on other types of tasks. On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's response contains an equals sign, indicating symbolic operations and reasoning. Following this finding, we analyze the behavior of CoT on these problems by separating planning and execution and comparing against tool-augmented LLMs. Much of CoT's gain comes from improving symbolic execution, but it underperforms relative to using a symbolic solver. Our results indicate that CoT can be applied selectively, maintaining performance while saving inference costs. Furthermore, they suggest a need to move beyond prompt-based CoT to new paradigms that better leverage intermediate computation across the whole range of LLM applications.
Forward citations
Cited by 44 Pith papers
-
Rationales Are Not Silver Bullets: Measuring the Impact of Rationales on Model Performance and Reliability
Rationale-augmented finetuning can hurt accuracy while improving calibration, with the sizes of both effects tied linearly to task difficulty.
-
Multimodal LLMs Can Reason about Aesthetics in Zero-Shot
A zero-shot two-stage prompting baseline (ArtCoT) makes multimodal LLMs' aesthetic judgments align substantially better with human expert rankings in pairwise artwork comparisons.
-
Chumor 2.0: Towards Benchmarking Chinese Humor Understanding
Chumor is a new Chinese humor explanation benchmark where LLMs perform near chance (best accuracy 60.3%) and well below human accuracy (78.3%).
-
LA-RL: Label-Aware Self-Reflection for Reinforcement Learning in Information Extraction
Label-aware diagnostic reflection plus two-stage outcome GRPO improves same-backbone IE F1 over SFT, with larger gains under relation-extraction domain shift.
-
UNIBROWSE: A Data-to-Agent Framework for Multimodal BrowseComp
A unified KG-plus-live-web data pipeline covering all three multimodal BrowseComp information-flow patterns, plus an exploration-degree filter, yields a 35B agent at 54.4 avg accuracy.
-
Simplicity Paradox: Debunking myths about prompting and datasets for LLM evaluation
In a 430k-evaluation study, plain baseline prompting beats most elaborate prompting techniques on non-reasoning LLMs across MCQA benchmarks, with only small role-framing variants gaining about 3 percentage points.
-
ScheMatiQ: From Research Question to Structured Data through Interactive Schema Discovery
ScheMatiQ uses LLMs to automatically generate schemas and extract structured data from text corpora based on natural language questions, supported by interactive user steering.
-
CoSPlan: Corrective Sequential Planning via Scene Graph Incremental Updates
VLMs perform near chance at detecting and correcting an erroneous step in visual sequential planning, and incrementally updating scene graphs step-by-step (SGI) yields small, partially inconsistent gains.
-
Rethinking Reasoning Quality in Large Language Models through Enhanced Chain-of-Thought via RL
DRER rewards CoT trajectories that increase the model's likelihood of the correct answer, plus a length penalty, and the new LogicTree benchmark reportedly lifts a 7B model's average accuracy from 0.13 to 0.60.
-
Dissecting Clinical Reasoning in Language Models: A Comparative Study of Prompts and Model Adaptation Strategies
Prompt structure and LoRA adaptation both strongly affect F1 on NLI4CT clinical NLI, but the claimed consistent +8 to 12 point LoRA gains and over 97% validity are not supported by the reported per-configuration results.
-
Retrieval-Confused Generation is a Good Defender for Privacy Violation Attack of Large Language Models
A retrieval-confused generation defense that swaps query comments with the most irrelevant paraphrased comments, reducing privacy attack success rates across eight LLMs.
-
MotiveBench: How Far Are We From Human-Like Motivational Reasoning in Large Language Models?
MotiveBench shows that large language models still lag human consensus on motivational reasoning, with GPT-4o scoring 80.89 percent and chain-of-thought prompting usually reducing accuracy.
-
Token Signature: Predicting Chain-of-Thought Gains with Token Decoding Feature in Large Language Models
The monotonicity of token probabilities during initial decoding predicts chain-of-thought gains, enabling dynamic selection between CoT and direct answers.
-
Around the World in 24 Hours: Probing LLM Knowledge of Time and Place
On GeoTemp, the best open model answers only 56% of two-city time questions and 33% when an hour shift is added, despite near-perfect scores on pure time arithmetic.
-
Fostering Video Reasoning via Next-Event Prediction
Next-event prediction, training video language models to caption unseen future frames, improves their scores on several temporal benchmarks while roughly preserving general video understanding.
-
Beyond path selection: Better LLMs for Scientific Information Extraction with MimicSFT and Relevance and Rule-induced(R$^2$)GRPO
MimicSFT plus R2GRPO improves scientific relation extraction in LLMs, beating supervised baselines and showing RLVR can expand reasoning capacity.
-
Are Large Language Models Reliable AI Scientists? Assessing Reverse-Engineering of Black-Box Systems
LLMs struggle to use passive observations for reverse engineering, but active intervention improves performance, largely through the process of generating queries rather than the data obtained.
-
Training Strategies for Efficient Embodied Reasoning
Reasoning pre-training and dropout let robot policies benefit from chain-of-thought training without run-time reasoning, preserving most of the performance gains at standard VLA inference speeds.
-
THOUGHTTERMINATOR: Benchmarking, Calibrating, and Mitigating Overthinking in Reasoning Models
Reasoning models are poorly calibrated to problem difficulty, overthinking easy questions; a new interrupt-based decoding method sharply reduces token spend with mostly stable accuracy.
-
MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency
A new multimodal benchmark evaluates chain-of-thought reasoning quality, robustness, and efficiency in large vision-language models, finding CoT often hurts perception tasks and that many reflection steps are unhelpful.
-
CoT-Valve: Length-Compressible Chain-of-Thought Tuning
A single LoRA task vector, scaled up or down at inference, controls chain-of-thought length in LLMs and compresses reasoning tokens with little accuracy loss.
-
When One LLM Drools, Multi-LLM Collaboration Rules
A position paper that introduces a four-level taxonomy of multi-LLM collaboration (API, text, logit, weight) and argues it is essential for reliability, pluralism, and democratization.
-
Retrieval-Based Interleaved Visual Chain-of-Thought in Real-World Driving Scenarios
A retrieval-based interleaved visual chain-of-thought method, RIV-CoT, improves VLM answer accuracy by 3.1 points and reasoning accuracy by 4.6 points on a new driving theory VQA benchmark.
-
INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge
INCLUDE is a multilingual benchmark of 197,243 exam questions from local sources that evaluates how well LLMs handle regional and cultural knowledge.
-
Beyond Examples: High-level Automated Reasoning Paradigm in In-Context Learning via MCTS
HiAR-ICL builds reusable 'thought card' reasoning templates via MCTS and shows a 7B model with these templates outperforms GPT-4o on MATH and AMC.
-
TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation
TrackRec trains a small LLM to generate user-preference summaries and a validator to score them, alternating the training so each improves the other, and reports gains on public and industrial recommendation benchmarks.
-
"Lost-in-the-Later": Framework for Quantifying Contextual Grounding in Large Language Models
LLMs ground answers in early context far more than later context, and chain-of-thought prompting or reasoning models reduce contextual grounding rather than improving it.
-
Hallucination Detection with Small Language Models
A multi-small-model ensemble with sentence splitting, z-score normalization, and harmonic mean detects hallucinations in RAG answers with a reported 10% F1 gain over single-model baselines.
-
SimuPanel: A Novel Immersive Multi-Agent System to Simulate Interactive Expert Panel Discussion
A multi-agent LLM system called SimuPanel simulates expert panel discussions with personas grounded in public academic sources, and a small evaluation suggests the full reasoning pipeline produces higher LLM-judged di...
-
Knowing Before Saying: LLM Representations Encode Information About Chain-of-Thought Success Before Completion
LLM hidden states encode enough information to predict chain-of-thought success before any reasoning tokens are generated, outperforming a text-only classifier.
-
SoftCoT++: Test-Time Scaling with Soft Chain-of-Thought Reasoning
SoftCoT++ generates multiple diverse latent thoughts through distinct initial tokens and a contrastive loss, improving reasoning accuracy beyond soft-thought and self-consistency baselines.
-
When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs
Chain-of-thought prompting degrades instruction-following accuracy on IFEval and ComplexBench across most tested models, and selective reasoning can partially recover the loss.
-
GenCLS++: Pushing the Boundaries of Generative Classification in LLMs Through Comprehensive SFT and RL Studies Across Diverse Datasets
SFT plus RL with post-hoc selection of the best training and inference prompts improves generative text classification accuracy by about 3.5% relative to a naive SFT baseline.
-
Perspective Transition of Large Language Models for Solving Subjective Tasks
Reasoning through Perspective Transition (RPT) improves LLM performance on subjective NLP tasks by ranking direct, role, and third-person perspectives by self-reported confidence and answering from the top-ranked perspective.
-
Disentangling Memory and Reasoning Ability in Large Language Models
Special tokens that separate memory recall from reasoning steps improve accuracy and interpretability on three QA benchmarks, but the gains over planning-token baselines are small.
-
Blueprint First, Model Second: A Framework for Deterministic LLM Workflow
A source-code-driven 'blueprint first' agent framework reports strong τ-bench results, but the paper contains an abstract/body mismatch and releases no artifacts.
-
Chain-of-Thought for Autonomous Driving: A Comprehensive Survey and Future Prospects
A survey that classifies chain-of-thought methods for autonomous driving into modular, logical, and reflective pipelines, and proposes three evolutionary stages from direct prompting to reinforcement learning.
-
Generative to Agentic AI: Survey, Conceptualization, and Challenges
Agentic AI is characterized over Generative AI by iterative reasoning, environment interaction, memory, and tool use, with autonomy as the defining difference.
-
Meta-Thinking in LLMs via Multi-Agent Reinforcement Learning: A Survey
The paper surveys existing work on LLM meta-thinking and argues that multi-agent reinforcement learning is a promising missing ingredient for building self-correcting language models.
-
Evaluation of LLMs for mathematical problem solving
A three-model, three-dataset LLM math evaluation using a multi-dimensional reasoning rubric, undermined by contradictory accuracy tables.
-
Large Language Models for Planning: A Comprehensive and Systematic Survey
A structured survey of LLM planning methods, benchmarks, and interpretability work, organized around a three-way taxonomy.
-
CoTGuard: Using Chain-of-Thought Triggering for Copyright Protection in Multi-Agent LLM Systems
A trigger-based watermark for multi-agent reasoning traces detects only the injected phrase, not the reproduction of copyrighted content.
-
Nature's Insight: A Novel Framework and Comprehensive Analysis of Agentic Reasoning Through the Lens of Neuroscience
A survey and taxonomy that organizes AI agentic reasoning into four neuroscience-inspired categories without introducing new empirical results.
-
A Survey on Large Language Models for Mathematical Reasoning
Recent advances in LLM mathematical reasoning are organized into comprehension and generation phases, covering methods from prompting to test-time scaling.
Discussion (0). Continue with ORCID to comment.