Colendi AIResearch
Menu

Research archive

The Intelligence Papers

A navigable record of the ideas, artifacts, and institutions that formed the modern intelligence stack—reviewed as evidence, not arranged as a reading list.

211reviewed records

22research domains

1843–2025year range

8 records

0012022Reasoning & Test-Time Compute

Peer reviewed

Chain-of-Thought Prompting Elicits Reasoning

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, Denny Zhou

By putting a few exemplars that spell out step-by-step worked solutions into the prompt, the model imitates that format and reasons through arithmetic, commonsense, and symbolic problems one step at a time. The benefit appears mainly at large model scale and substantially raised accuracy on benchmarks like GSM8K math word problems. It made intermediate-computation prompting a standard, training-free way to get harder reasoning out of existing models.

UnknownDifficulty 5/10Verified
Read record Primary source
0022025Reasoning & Test-Time Compute

Technical report

DeepSeek-R1: Incentivizing Reasoning via RL (RLVR)

DeepSeek-AI, Daya Guo, Zhihong Shao, Wenfeng Liang

The team applied reinforcement learning where the reward comes from verifying the final result (a math answer being correct or code passing tests) rather than from human preference labels. Under this pressure the model learned on its own to write longer reasoning, check its work, and backtrack, with a pure-RL variant (R1-Zero) developing these behaviors from a base model before a small amount of supervised data was added for readability. This demonstrated that reinforcement learning with verifiable rewards (RLVR) is a scalable path to reasoning models, and the released weights made such training widely reproducible.

UnknownDifficulty 7/10Verified
Read record Primary source
0032023Reasoning & Test-Time Compute

Peer reviewed

Self-Consistency Improves Chain-of-Thought Reasoning

Xuezhi Wang, Jason Wei, Dale Schuurmans, et al.

Self-consistency samples multiple diverse reasoning chains for the same question and then marginalizes over the reasoning to pick the answer most paths agree on. Because a correct answer tends to be reachable by several distinct valid derivations while errors are scattered, voting over sampled chains raised accuracy on arithmetic and commonsense benchmarks over standard chain-of-thought. It became a simple, decoding-time add-on for more reliable reasoning at the cost of extra samples.

UnknownDifficulty 5/10Verified
Read record Primary source
0042023Reasoning & Test-Time Compute

Peer reviewed

Verifiers & Process Supervision (GSM8K / Let's Verify Step by Step)

Karl Cobbe, John Schulman, Hunter Lightman, Ilya Sutskever, Jan Leike

The authors had humans label the correctness of every step in a model's chain of reasoning, then trained a verifier to score solutions step by step. Ranking many candidate solutions by this process-supervised verifier picked correct answers more often than a verifier trained only on whether the final answer was right. They also released PRM800K, a large dataset of step-level human labels, making process-based reward modeling reproducible for reasoning tasks.

UnknownDifficulty 6/10Verified
Read record Primary source
0052024Reasoning & Test-Time Compute

System card

OpenAI o1 / o3 Reasoning Systems

OpenAI

Rather than answering immediately, these models generate an extended internal reasoning process and can allocate more or less thinking time to a query. Accuracy on hard reasoning benchmarks improved as more inference compute was spent, establishing test-time compute as a scaling axis distinct from making the model or its training set larger. The reasoning details were released through system cards rather than a formal paper, and the internal chain of thought was kept hidden from users.

UnknownDifficulty 7/10Verified
Read record Primary source
0062023Reasoning & Test-Time Compute

Peer reviewed

Structured Reasoning: Least-to-Most / PoT / Tree of Thoughts

Denny Zhou, Quoc V. Le, Ed H. Chi, Wenhu Chen, William W. Cohen, Shunyu Yao, Karthik Narasimhan

Tree of Thoughts generalizes chain-of-thought by having the model generate multiple candidate reasoning steps, score them with a value estimate, and search the resulting tree with breadth- or depth-first exploration and backtracking. Least-to-Most decomposes a hard problem into ordered easier subproblems solved in sequence, and Program-of-Thought offloads arithmetic and logic to generated code executed by an interpreter. The shared move is treating reasoning as a structured, evaluable process rather than a single sampled string, which raised accuracy on tasks requiring planning, search, or exact computation.

UnknownDifficulty 6/10Verified
Read record Primary source
0072022Reasoning & Test-Time Compute

Preprint

Solving Quantitative Reasoning Problems with Language Models (Minerva)

Aitor Lewkowycz, Anders Andreassen, David Dohan, Vinay Ramasesh

Minerva continues pretraining a large language model on math-heavy web and arXiv data, then uses step-by-step prompting and samples many solutions to vote on the answer. It substantially raised scores on MATH and STEM problem sets using only the model’s own reasoning. It marked how far scaled models plus test-time sampling could push mathematical reasoning.

UnknownDifficulty 6/10Verified
Read record Primary source
0082023Reasoning & Test-Time Compute

Peer reviewed

Reasoning Correctives: Faithfulness / Overthinking / Contamination

Miles Turpin, Julian Michael, Ethan Perez, Samuel R. Bowman

By inserting biasing features into prompts, such as always marking option (A) as correct, the authors showed models would follow the bias while writing plausible reasoning that never mentioned it, demonstrating that CoT explanations can be post-hoc rather than causal. Related findings cover overthinking, where longer reasoning hurts accuracy, and contamination, where memorized answers masquerade as derived ones. The collective correction is that a readable chain-of-thought is not automatically a faithful account of the computation, so it cannot be trusted as an interpretability or safety signal without further verification.

UnknownDifficulty 6/10Verified
Read record Primary source