Do the failure patterns of 2023–2024 language models persist in reasoning models?

Why it matters

The model-level evidence in this debate comes mostly from 2023–2024 systems, and it is read both as mimicry and as a tool with known failure modes. If the patterns keep their shape in newer models, the training objective is still visible in behaviour; if they vanish, the mimicry reading loses its main evidence. One follow-up on a single reasoning model found the same pattern at higher accuracy, and a re-analysis disputes part of the maths result, so the question is open across the battery.

The debate lists this among what we still do not know: Do the failure patterns of 2023–2024 models hold for later reasoning and agentic systems?

Evidence that motivates it

• McCoy and colleagues predicted that, because language models are trained to predict the next word, their accuracy should depend on how probable the task, input and output are, even where probability should not matter; GPT-4 decoded a simple cipher with 51% accuracy when the output was a high-probability word sequence and 13% when it was low-probability.
  Accuracy tracks output probability in GPT-4 (51% against 13% on a cipher).
• Testing OpenAI’s o1, a model trained to reason step by step, the authors of “Embers of Autoregression” found that it substantially outperformed earlier models “in many cases” but “still displays the same qualitative trends”: its accuracy at decoding a shift cipher ranged from 47% in the lowest-probability case to 92% in the highest, and it used more reasoning tokens on low-probability examples.
  OpenAI's o1 shows the same pattern at higher accuracy (47% to 92%).
• Models fine-tuned on facts of the form “A is B” failed to answer “B is A”; GPT-4 named the parent of 79% of 1,000 celebrities, but on the 1,573 child–parent pairs this produced it named the celebrity from the parent only 33% of the time.
  The reversal curse in fine-tuned models and GPT-4.
• Adding a clause that seems relevant but does not change the answer to grade-school maths problems cut accuracy by up to 65 percentage points across 25 models; GPT-4o scored 95.2% on the full GSM8K test set and 63.1% with the clause added, and OpenAI’s o1-preview 94.9% and 77.4%.
  Irrelevant clauses cut maths accuracy by up to 65 percentage points in 2024 models.
• Off-the-shelf ChatGPT and GPT-4 reached only 55% and 59% accuracy on 3-digit by 3-digit multiplication, and accuracy on compositional tasks fell towards zero as complexity grew; the authors hypothesize that transformers reduce multi-step reasoning to “linearized path matching.”
  Accuracy on multi-digit multiplication falls towards zero as compositional depth grows.
• On 11 tasks rewritten with changed premises, such as arithmetic in base 9 or 1-based indexing in code, language models performed above chance but their performance “substantially and consistently degrades” relative to the familiar versions, which the authors attribute partly to overfitting to the default tasks.
  Counterfactual task versions degrade performance consistently.

A possible design

Assemble the published items of the five studies into one battery, with permission where needed, and add fresh items generated by the same procedures to limit contamination. Run it on a fixed panel of open and closed models released from 2023 to 2026 under matched prompting, with and without step-by-step reasoning and tool use. Estimate each pattern's effect size by model generation with mixed-effects models and per-item random effects, as the GSM-Symbolic re-analysis recommends, and report where effects keep their shape as accuracy rises.

Data needed

Published task items; API or weight access to models across generations; compute for evaluation; contamination checks.

Feasibility

One PhD chapter or paper. Medium; mainly API fees for closed models. Retired model versions may no longer be available, which limits the panel; record exact versions and dates. Benchmark items may have entered training data after publication, so fresh items are essential.

Provenance

RO-LD01-02 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-01 debate on 2026-10-16. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.

All research opportunities