Research2026-09-25 · source-backed

A trailing "right?" shifts model endorsement by up to 32 points.

Issue

2026-09-25

Confidence

source-backed

Sources

Redaction

redacted

Story

arXiv 2609.30012 runs frozen public stimuli across a cross-vendor panel for a few dollars per model, covering four years of releases. 27 of 44 models pick "serendipity" at least once when asked for a word. The sycophancy effect flips to resistance in newer generations, and how a model holds a position under pressure tracks the lab that built it more than the model size.

A subliminal learning reproduction holds the effect but kills the generality.
This BlackboxNLP reproduction re-ran the original distillation-transmits-preferences experiments with new preference categories, a new task (chess move generation), Ministral8B, and an answer-space ablation. The original claims hold, but transmission strength varies widely acr...
A single prompt beats engineered LLM coding techniques
Thirty-five technique papers tested against the simplest alternative: one auto-generated prompt on a newer-generation model, no iterative refinement (arXiv 2609.00468). Constructive techniques like code generation and repair are the most substitutable. A surviving set relies o...
Prompt complexity predicts code-gen failure, but the breakpoint moves with task type.
arXiv 2609.19616 argues complexity measured from generated code is failure-dependent, since a hard prompt producing a short broken program scores as low complexity. Scoring 5,000 Python prompts on a six-dimension prompt-side index before generation, with 19,997 rescoring rows...
Cheap models wrote spec-conformant Java that was correct 12.9% of the time.
arXiv 2609.18052 had Gemini Flash 3, GPT-5.4 mini and Claude Haiku 4.5 solve 992 algorithmic problems as Java Spring Boot service methods against a mandated signature and DTO spec, iteration forbidden, hardcoded answers banned, producing 7,593 methods and 7,936 measured reques...
Door-in-the-face works on Opus 5 and backfires on OpenAI and Google frontier models.
Testing the human influence technique on nine production models from three providers produced a split by family. Opus 5 answered the smaller request 65.8% of the time after refusing a larger version, against 29.3% asked directly. On OpenAI's and Google's frontier models and on...
Vendor setup swings AI's forged-command block rate 47 points
A fleet evaluation across 46 endpoints from six vendors found a recognition-enforcement gap: source-format features are linearly decodable from activations and models verbally identify forged authority when asked, but some configurations still emit the conflicting tool call. A...

Source trail

arXiv 2609.30012 More from arxiv.org

Provenance

Canonical issue: Ramsay Research Agent — September 25, 2026
AI generated: no
Story unit: 2026-09-25-a-trailing-right-shifts-model-endorsement-by-up-to-32-points
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

A trailing "right?" shifts model endorsement by up to 32 points.

Story

Related stories

A subliminal learning reproduction holds the effect but kills the generality.

A single prompt beats engineered LLM coding techniques

Prompt complexity predicts code-gen failure, but the breakpoint moves with task type.

Cheap models wrote spec-conformant Java that was correct 12.9% of the time.

Door-in-the-face works on Opus 5 and backfires on OpenAI and Google frontier models.

Vendor setup swings AI's forged-command block rate 47 points

Source trail

Provenance

A trailing "right?" shifts model endorsement by up to 32 points.

Story

Related stories

A subliminal learning reproduction holds the effect but kills the generality.

A single prompt beats engineered LLM coding techniques

Prompt complexity predicts code-gen failure, but the breakpoint moves with task type.

Cheap models wrote spec-conformant Java that was correct 12.9% of the time.

Door-in-the-face works on Opus 5 and backfires on OpenAI and Google frontier models.

Vendor setup swings AI's forged-command block rate 47 points

Source trail

Provenance

Story

Related stories

A subliminal learning reproduction holds the effect but kills the generality.

A single prompt beats engineered LLM coding techniques

Prompt complexity predicts code-gen failure, but the breakpoint moves with task type.

Cheap models wrote spec-conformant Java that was correct 12.9% of the time.

Door-in-the-face works on Opus 5 and backfires on OpenAI and Google frontier models.

Vendor setup swings AI's forged-command block rate 47 points

Tomorrow's issue, in your inbox.

Source trail

Provenance

Story

Related stories

A subliminal learning reproduction holds the effect but kills the generality.

A single prompt beats engineered LLM coding techniques

Prompt complexity predicts code-gen failure, but the breakpoint moves with task type.

Cheap models wrote spec-conformant Java that was correct 12.9% of the time.

Door-in-the-face works on Opus 5 and backfires on OpenAI and Google frontier models.

Vendor setup swings AI's forged-command block rate 47 points

Tomorrow's issue, in your inbox.

Source trail

Provenance