Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

counterfactual reasoning

Counterfactual reasoning is the mental act of asking what would have happened if some action, event, or condition had been different. It lets you compare an actual outcome with an imagined alternative, such as “if we had used another drug, would the patient have recovered faster?” or “if I had turned left instead of right, would I have avoided the traffic jam?” It is a tool for understanding cause and effect, planning, and assigning responsibility, because it separates what happened from what might have been.

The papers show counterfactual thinking is powerful but not foolproof. It helps models estimate unobserved memory values, guide reinforcement learning agents, and detect bias in reasoning, yet it also reveals that models can create new fairness violations, struggle to predict outcomes for unseen situations, and need intermediate feedback to answer accurately. A recurring lesson is that asking “what if” improves performance, but the answers are only as trustworthy as the assumptions and context baked into the question.

Papers
109
Released code
3
First seen
Aug 2026
Latest
Nov 2026

106 papers in the last two months, against 3 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1AI agents give plausible explanations but fail to recover scientific laws

    Current AI agents can investigate systems actively but cannot yet induce exact scientific laws in unfamiliar settings.

    Oct 2026
  2. 2Language models can resist distractors better without additional training

    Weight editing suppresses spurious reasoning in LLMs, improving accuracy on perturbed inputs by 11.8 percentage points.

    Nov 2026
  3. 3AI models can forecast sales well yet misjudge price changes

    Forecast accuracy does not indicate how accurately a model predicts the effect of price changes.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

credit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.15 togethermechanistic interpretabilityThe study of how neural networks compute through internal circuits, features, and causal interventions, rather than just input-output behavior.15 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.14 together

Who is working on it

Alibaba Group 5University of Wisconsin–Madison 4Stanford University 4Independent Researcher 4Tencent 4Seoul National University 3

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.