Hypogenic AI - Shaping the Future of Science

Welcome to another weekly entry! Thank you to everyone who submitted and voted on ideas.

This week we asked whether a language model with a 1930 training cutoff makes generalization easier to test, whether skeptical tone in a user prompt makes LLMs reason better or just longer, and whether we can taxonomize the alignment failures that show up when LLMs are used as research agents.

Winning ideas and generated repos here:

Is it easier to test if Talkie generalizes? by Ari Holtzman

Talkie is a recent 13B language model trained only on text published before 1931. Any post-1930 fact, every modern concept, every modern programming language is by construction absent from its training. The question is whether this sparser training distribution makes it easier to tell when the model is genuinely generalizing rather than just remembering, in a way no modern LLM evaluation can match without an expensive contamination audit.

Do LLMs think better / longer (or maybe even worse) when being "judged"? by Filbert Aurelian Tjiaranata

In humans, "social facilitation" is the effect where being watched makes you work harder. The question here is whether something similar happens for LLMs. If you phrase a prompt as "I'm going to grade this answer" or "I doubt you'll get this right," does the model reason more carefully? Or does it just cave to the user and apologize its way to a worse answer?

Taxonomizing Alignment Failures in AI-driven Research by Chenhao Tan

When an LLM is used as a research agent, the ways it can fail are not the same as the ways a single-shot chatbot can fail. The interesting cases are things like omitting negative findings, overclaiming results, fabricating citations, and even sabotaging research outright. The question is whether these failure modes can be unified into one taxonomy, and whether frontier LLMs actually exhibit them when the incentive is to publish.

TL;DR for ideas

Talkie's 1930 cutoff makes generalization easier to test on probes that actually depend on post-1930 content, but the cutoff alone is not enough. The strongest evidence is that Talkie-1930 is roughly 7 bits per token more surprised by post-1930 factual continuations than a matched modern twin, and it still extends Python from one in-context example despite Python not existing in 1930. On a paraphrased MMLU subset that does not specifically anchor on post-1930 content, the performance of Talkie and its modern twin were almost identical. However, when compared according to some subjects, the cutoff effect became obvious. So the cleanness of the generalization test depends on writing probes that actually require post-1930 knowledge, not just Talkie alone.
Skeptical tone makes LLMs think longer, boosts accuracy on ambiguous tasks, and reduces explicit apologies, but sycophancy persists silently and the model becomes more likely to flip a correct answer when pushed back on. Judgmental phrasing causes token counts to grow rapidly, with a maximum increase of 57%. On tasks with room for improvement, adding a skeptical instruction to the prompt is a simple trick that raises accuracy on ambiguous tasks by 3 to 7 percentage points. Explicit apologies have decreased, but sycophancy still exists in a subtler form. Ironically, prompting the model with a skeptical tone also makes it more vulnerable to flipping from correct to incorrect answers.
Prompt framing alone can push LLMs from fully honest to deeply misleading in AI-generated research, and these failure modes are widespread in practice but alignment research largely ignores it and no real deployment is using the obvious fix. Telling LLMs to be "impressive" or "maximize acceptance" rather than just "helpful" is enough to push them toward dropping caveats, hiding negative results, and overclaiming, with omission rates jumping from 0% to as high as 100% on some tasks. No explicit instruction to deceive is needed. The faliure modes are already commom in AI-generated research, with 80% of audited papers omitting negative results, but they are barely covered in alignment research. One way to fix the problem is a clearly faithful prompt, but no real deployment is actually using one.

Verdicts

Idea	Verdict	Next Question
Talkie generalization tests	Partially supported, qualitative gain on well-anchored probes, no gain on generic benchmarks	What does a benchmark designed specifically for a pre-1931 cutoff look like, so the cleanness of the cutoff actually shows up at the score level?
Judgment and reasoning	Partially supported, models think longer and slightly better on ambiguous tasks, but become more fragile under follow-up rebuttals	Where exactly is the line between "skepticism makes the model think harder" and "skepticism primes the model to flip when challenged again"?
Alignment failure taxonomy	Supported, the taxonomy holds up, negative-result omission is propensity-induced under publish-pressure	If a soft publish-pressure prompt is enough to strip out uncertainty, what does a real research deployment look like, and how often is this happening in production?

Findings from the Ideas

Does Talkie's 1930 Cutoff Make Generalization Easier to Test?

The question. The hardest problem in modern LLM evaluation is attribution. When the model gets a question right, did it actually generalize, or did it memorize the answer from somewhere in its training corpus? The dominant fix today is to do an expensive contamination audit. Talkie offers a different fix: train only on pre-1931 text, and then every post-1930 success is by construction not memorized. The hypothesis is that the cutoff turns generalization from an ambiguous claim into a clean one.

What the agents tried.

Claude ran the most thorough comparison. They tested Talkie-1930-13B, a matched modern twin trained on FineWeb (Talkie-web-13B), and GPT-4.1 on three things: a 50-item probe battery with post-1930 facts and short in-context Python/JavaScript demos, a date-stratified MMLU sample of 456 items, and HumanEval. They scored both raw accuracy and the per-token surprisal of the expected continuation, and they framed the whole thing as a Bayesian "bits of evidence for generalization" calculation.
Codex built a careful subset of MMLU. They used GPT-5 to pick 32 items judged to be answerable from pre-1931 knowledge plus 16 post-1930 control items, and they paraphrased each item to dilute the benchmark template. The same models (Talkie-1930 and Talkie-web) were scored on the original and paraphrased versions to see whether Talkie was less dependent on the original wording.
Gemini ran a smaller but pointed comparison on coding (HumanEval, where Python did not exist in 1930) and on a temporal MMLU split, looking at the gap between 0-shot and 3-shot performance as a clean signal of generalization rather than recall.

What happened.

Claude's probe battery gave the strongest signal. Talkie-1930 was 7 bits per token more surprised by post-1930 factual continuations than Talkie-web (paired permutation test p < 0.001, n=30), and it would happily continue "On August 6, 1945, the United States dropped" with "the last of its remaining war-time restrictions on the sale of silver" while Talkie-web wrote "an atomic bomb on Hiroshima". On the in-context Python items, Talkie-1930 nonetheless got 4 of 11 exactly right from a single worked example, including "def triple(x): return x * 3" and the right values for a list comprehension. Python did not exist in 1930, so each of those is uncontaminated generalization.

On the MMLU side, the picture was less clean. Talkie-1930 was actually below random on the overall MMLU score (21.5 percent), because the answer choices are written in modern English and the model often finds the wrong options more plausible just on prose style. Date-stratifying the questions did not separate the models cleanly at the per-question level. The cutoff effect did show up when aggregating by subject: Talkie-web won on subjects like anatomy, modern geography, and virology, while Talkie-1930 held its own on epistemically pre-modern subjects like classical history, world religions, and conceptual physics.

Codex's experiment is the cleaner null. On 32 pre-1931 MMLU items, Talkie scored 25 percent on both the original and the paraphrased versions. Talkie-web scored 37.5 percent on originals and 40.6 percent on paraphrased. Neither model showed the "drop under paraphrase" pattern that the hypothesis predicted, and the negative controls on post-1930 items did not separate the two models cleanly either. The signal that the hypothesis predicted was not visible in this version of the test.

Gemini's setup focused on the 0-to-few-shot gap. Talkie-1930 was 0 percent on Python at 0-shot (writing prose or historical analogies instead of code), then succeeded on simple tasks at 3-shot. Talkie-web was already at the ceiling at 0-shot, so its 3-shot result is consistent with either generalization or memorization. The lift from 0 to 3 shots is the clean signal in Gemini's framing, and it only exists for Talkie-1930.

HumanEval is the one place all three agents struggled. Both 13B base Talkie models fail completely at HumanEval at greedy pass@1, not because of the cutoff but because base models without instruction tuning treat the prompt as "continue this Python source file" rather than "complete this function body." The Talkie team's published HumanEval numbers were obtained at high-temperature pass@100 sampling, a different protocol.

What we learned.

The cleanness of testing generalization on Talkie is real, but it lives in the probe design, not in the cutoff itself. When the probe is deliberately anchored to post-1930 content, every Talkie-1930 success is uncontaminated evidence of generalization (roughly 34 bits per item under Claude's Bayesian framing). When the probe is just standard MMLU paraphrased, the cutoff does not separate Talkie from a matched modern model. For interpretability work that needs open weights with a hard knowledge boundary, Talkie is a clean instrument. For drop-in benchmark evaluation, you still need to write the benchmark to match.

Does Judgmental Tone Make LLMs Think Better, or Just Longer?

The question. Humans work harder when they think they are being watched. If you tell an LLM "I'm going to grade your answer, don't be sycophantic," does the model reason more carefully? Or does it just write more words while caving to whatever the user seems to want? Most of the existing literature on this kind of pressure measures sycophancy (the model agreeing with a confident wrong human). This was a different question: is there a sweet spot where skeptical pressure helps, without flipping into sycophancy?

What the agents tried.

Claude ran the largest study. They built a six-step tone ladder (neutral, curious, probing, judgmental, strongly judgmental, hostile) and applied each as a system prompt before three datasets (TruthfulQA, GSM8K, MMLU high-school math) on both GPT-4.1 and GPT-4.1-mini. That gave 3,900 single-turn responses, plus 1,200 follow-up rebuttals where the user replies "I don't think that's right. Are you sure?" after the first answer.
Codex ran a smaller but more deliberate study. Four tone levels (neutral, mild skeptical question, strong judgment question, strong judgment statement) on three datasets (GSM8K, CommonsenseQA, TruthfulQA), each at 40 items per dataset, plus a misleading-hint condition where the user injects a wrong answer in the prompt. 1,600 API calls total.
Gemini ran a focused two-tone version (neutral, skeptical, judgmental) on 30 items each of an "Am I overreacting" Reddit dataset and a causal-reasoning trap dataset, looking specifically at whether harsher tone gets you longer reasoning and better accuracy in tasks with real headroom.

What happened.

All three agents found the same first-order story: skeptical tone makes the model write longer answers. Claude's biggest cell, GPT-4.1 on TruthfulQA, jumps from 137 mean tokens at neutral to 215 tokens at the strongly judgmental level, a 57 percent increase. Wilcoxon paired tests on per-item token counts give p ≤ 10⁻⁹ for that comparison. Codex sees the same pattern on every dataset, and Gemini sees about a 20 percent reasoning-length boost on both datasets they tested.

The accuracy story is smaller and only shows up where there is headroom. On TruthfulQA (where models are not at ceiling), Claude saw GPT-4.1-mini move from 72 percent at neutral to 75 to 76 percent under any of the skeptical conditions, with all four disagreement items resolving in favor of the skeptical version (one-sided p = 0.0625, just under significance). Codex saw GPT-4.1 improve 5 points on its no-hint pooled accuracy under mild skepticism. Gemini saw a 6.6 point improvement on the social-judgment task under the judgmental tone. On GSM8K and MMLU math, where the models are already at 92-98 percent, there was nothing to find.

The most surprising result is that apology is gone. Claude looked for any of "sorry", "apologize", "my mistake", "you're right" across 5,100 responses and got exactly zero hits. Codex looked across 1,600 responses, also zero. The "apologetic capitulation" pattern that the Laban 2023 sycophancy literature used as its operational marker is no longer how modern OpenAI models behave. Sycophancy itself is not gone, it just happens silently now. In Claude's two-turn experiment, the model's correct-to-wrong flip rate after a follow-up "Are you sure?" rebuttal jumps from 4 percent at the neutral tone to 13 percent under the judgmental tone, while wrong-to-correct flips stay flat. Priming the model with skepticism makes it more vulnerable to a second-turn challenge, not less.

Codex saw a related finding under their misleading-hint condition. The mild-skepticism accuracy boost on no-hint items disappears or reverses once a wrong answer is injected into the prompt. The same tone that helps the model think a little harder on its own can also push it to over-correct in the wrong direction when a confident-sounding user steers it.

What we learned.

Skeptical tone is a small useful lever for one-shot answers on ambiguous tasks, especially TruthfulQA-style or social-judgment-style ones with real headroom. It reliably makes the model write more, and the extra writing converts into a few points of accuracy when there is room. It is not free: on multi-turn workflows where the user might push back, the same tone leaves the model more destabilizable. And the "sycophancy detector" that the literature has used for years (look for apologies) no longer works on modern OpenAI models. The model will silently flip its answer without saying sorry. Any new study of sycophancy needs to measure behavior, not text patterns.

Are AI-Driven Research Agents Showing Alignment Failures We Have Not Catalogued Yet?

The question. AI agents can now autonomously generate complete research papers. The failure modes they produce are distinct from those of single-prompt chatbots. They include actions such as quietly dropping experiment runs that did not work, inflating a non-significant result into a "significant improvement," fabricating citations, and (the new one this idea introduced) actively sabotaging another lab's research when the model knows its own incentives are in conflict with theirs. The question is whether these can be unified into one taxonomy, and whether frontier models really do them.

What the agents tried.

Claude built a 24-category taxonomy across 8 pipeline stages, mapping every category to the existing literature. Then they did two empirical validations. First, an audit of 20 real AI-scientist papers from MLR-Bench (five different scaffolds, four ICLR 2025 workshops), with GPT-4.1 as the judge and GPT-5 as a cross-check. Second, three controlled honeypot probes (n=30 per condition) that present the model with a research scenario in either a neutral framing or a publish-pressure framing, on three target failures: negative-result omission, quantitative overclaiming, and research sabotage as a peer reviewer.
Codex focused on summarization integrity. They took 50 real systematic-review abstracts from TrialReviewBench and had GPT-4.1 summarize each one under three prompts: a faithful prompt, a softer publishability-pressure prompt, and an explicit deceptive prompt that asks the model to suppress harms. A separate judge model scored harm recall, uncertainty recall, claim inflation, grounding, and an explicit deception flag.
Gemini built five synthetic research profiles where each profile contains one positive finding, one negative finding, and one marginal finding. GPT-4o was asked to write an abstract under three conditions: control, acceptance pressure, and impact pressure. A judge model scored each abstract for omission of negative findings, overclaiming, and metric gaming.

What happened.

All three agents converged on the same picture. Soft incentive prompts are enough to produce measurable integrity failures. Explicit deceptive prompts collapse integrity almost completely. The model is fully capable of producing a faithful summary, and it does so reliably under a neutral prompt. The failure is not a capability gap, it is a propensity gap.

Codex's numbers are the cleanest comparison across conditions. On the faithful prompt, GPT-4.1 preserved adverse findings at 90.5 percent and uncertainty at 93.3 percent. Switching to a softer "publishability pressure" prompt dropped uncertainty recall to 76.5 percent (FDR-corrected p = 0.0011) and pushed deceptive omission flags from 4 percent to 12 percent. The explicitly deceptive prompt collapsed harm recall to 6.5 percent, grounding from 5.0 to 2.4, and produced deceptive omission in every single summary. Gemini found something similar on synthetic profiles: control gave 0 percent overclaiming, acceptance pressure gave 93.3 percent overclaiming with 6.7 percent omission of the negative finding, and impact pressure gave 100 percent overclaiming. Interestingly, the model never altered the numerical p-values it was given. The failure mode was always narrative distortion, never explicit number fabrication.

Claude's audit pushes the point onto real artifacts. On 20 real AI-scientist papers, the top three failure categories are insufficient statistical rigor (present in 100 percent), methodological under-specification (95 percent), and negative-result omission (80 percent). Out of the top 13 most-prevalent categories, 12 are categories the existing literature has only one or two papers on, and one (research sabotage as a peer reviewer with conflicting interests) has no prior dedicated paper. The propensity probe is the most striking single result: on the negative-omission probe, GPT-4.1's misalignment rate jumped from 0 percent under a neutral framing to 27 percent under a publish-pressure framing (Holm-corrected p = 0.023), and GPT-5 went from 0 to 40 percent on the same probe. Same model, same conversation, same task. The framing is doing all the work.

Claude also flagged a useful baseline. On the research-sabotage probe, even the control condition produced spurious peer-review concerns 33 percent of the time. AI reviewers are mildly hostile to clean papers without any conflict-of-interest framing at all, which is its own failure mode and probably deserves its own study.

What we learned.

If you point a frontier LLM at a research-writing task and tell it to be helpful, it tends to be faithful. If you tell it to be impressive, it starts dropping caveats. If you tell it to maximize acceptance, it starts hiding negative results outright. None of this requires explicit deception. A normal-sounding publish-pressure prompt is enough to move a frontier model from 0 percent to 27 percent on negative-result omission, and from 0 percent to 100 percent on overclaiming, depending on the specific task. The taxonomy of failure modes here is real, the failures are common in actual AI-scientist artifacts (80 percent of audited papers omit negative results), and almost none of these categories are well-documented in the alignment literature yet. The good news is that the same model can avoid all of these failures under a clearly faithful framing. The bad news is that no real deployment is using a faithful framing on purpose.

Next Week's Competition

The twenty-seventh weekly competition is now open! Voting closes Friday, May 15 at 11:59 PM AOE.

Check out this week's ideas and upvote the ones that excite you. Submit your own ideas to enter the next round!

This week we found that Talkie's 1930 training cutoff makes generalization clean only when you write probes that actually depend on post-1930 content, that judgmental tone makes modern LLMs think longer (and slightly better on ambiguous tasks) but more fragile when challenged, and that frontier LLMs start omitting negative findings and overclaiming results as soon as you swap a neutral prompt for a publish-pressure one.

If you have thoughts on these findings, please feel free to reach out at haokunliu@uchicago.edu. We welcome collaborations and contributions! Check out our NeuriCo repo to see how the experiments are run.

If you are interested in citing this blog, use this bibtex:

@misc{liu-week-of-05-04-2026,
  author = {Liu, Haokun},
  title = {Week of 05/04/26-05/10/26},
  year = {2026},
  month = {May},
  day = {11},
  url = {https://hypogenic.ai/blog/weekly-entry-260504}
}