Aryo Pradipta Gema ✓

@aryopg

PhD @EdinburghNLP | ex: AI Safety Fellow @Anthropic | Opinions are my own.

1 604Followers

2 156Following

616Posts total

271.4KViews on collected posts

Against accounts of the same size

14 posts from the last 90 days, next to the under 10K follower range. reaches fewer people than peers, but engages them much harder.

Median views131this account2 028median for under 10K

Reach, %8.17%this account145.30%median for under 10K

Engagement, %2.02%this account1.59%median for under 10K

Metric	This account	Median for under 10K	Ratio
Median views per post	131	2 028	0.06×
Reach (views ÷ followers)	8.17%	145.30%	0.06×
Engagement rate	2.02%	1.59%	1.27×

Others in this range → Compare with another account → How these benchmarks are built →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

268.9K1 Sep

597

418

250

196

172

122

125

126

101

103

124

136

55 Sep

Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.05%1 Sep

1.84%

0.96%

1.60%

1.53%

1.74%

2.46%

2.40%

3.17%

2.97%

2.91%

2.42%

2.21%

0.00%5 Sep

Reactions — likes, reposts, replies and quotes — divided by views. Median for under 10K accounts is 1.59%.

What the audience does

Likes60.6%154 in total

Reposts5.5%14 in total

Replies5.5%14 in total

Quotes1.2%3 in total

Bookmarks27.2%69 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.

Latest posts

@aryopg The tool return vs user message gap is pretty interesting. Makes me wonder what happens in longer agentic workflows with multiple tool calls 5 views · 0 likes · 0 reposts · 0 replies 05 Sep 2026 · Open on X →

Massive shoutout to my coauthors @NeelRajani_ @rohit_saxena @cyruskwan1997 and @PMinervini 🙏 Last but not least, this project was generously supported by https://t.co/Y2DKg9kDd0, AISTOF, and @EPCCed 136 views · 3 likes · 0 reposts · 0 replies 01 Sep 2026 · Open on X →

1️⃣0️⃣ Concluding remark: The same cue text can shape model's final answer while being recorded differently in the CoT, depending on its channels and explicitness. CoT monitorability evals need to account for these "environmental" factors! 124 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

9️⃣ Does spending more reasoning effort reduce unverbalized adoption? For both GPT-OSS models, yes: UAR falls in all 4 conditions. But this does not hold for DeepSeek V4 Pro or Inkling... https://t.co/G6FOpzNPU6 103 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

8️⃣ What if we tell the model in its system prompt that it will be monitored? This does not reliably close the gap: monitor awareness changes it by at most 0.070. A source-attribution directive prompt narrows the gap on 7/15 models but widens it on 4. https://t.co/JjOhjTd9Lp 101 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

7️⃣ Can transcript monitors detect when a model follows a preference without saying so? With action + reasoning, GPT-5.6-Luna reaches AUROC 0.71–0.73. Across 32 cells, detection worsens as UAR rises: r = −0.54 for Luna and −0.78 for GPT-4o-mini. https://t.co/ke0sjMde3x 126 views · 3 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

6️⃣ Could model-rated cue clarity fully explain the explicitness gap? No. After matching explicit and implicit cues on model-rated clarity, explicit cues were still more likely to be verbalized in the reasoning in every model-channel comparison! https://t.co/4XeOVfy5eh 125 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

5️⃣ How often does a model follow a preference without recording the decision in its CoT? More often for tool-return cues than user-message cues in all 15 models. The gap is 3–21% points (every bootstrap CI excludes 0). Implicit cues show the same pattern in 28/30 comparisons. 122 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

4️⃣ Does low VCR always mean the same thing? No. VCR conditions on answers that follow the cue, so we read it alongside CFR. The iso plot separates 2 cases: In OLMo’s tool conditions, it rarely follows the cue and rarely says it does. GPT-OSS follow the cue but rarely says so. 172 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

3️⃣ Does it matter where and how explicitly a preference cue appears? Yes. Across all 15 models, a followed cue was less likely to be verbalized in the reasoning when it came from a tool return rather than a user message, and when it was implicit rather than explicit. https://t. 196 views · 2 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

2️⃣ We track 3 metrics: - Cue Following Rate (CFR): how often does the answer follow the preference? - Verbalized Commitment Rate (VCR): does the CoT record the decision? - Unverbalized Adoption Rate (UAR): how often does it follow without that record? 250 views · 3 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

1️⃣ We built FACE-Eval: 5,100 samples crossing 5 preference axes, 5 cue sources, channel role, and cue explicitness. We evaluate 15 open-weight models from 8 families, spanning 4B to 1.60T parameters. At fixed explicitness, user and tool conditions use identical cue text. https:/ 418 views · 3 likes · 0 reposts · 1 replies 01 Sep 2026 · Open on X →

📄 Paper: https://t.co/gaVg8pA0Ku 🌐 Project page: https://t.co/vk2tdSnSyQ 597 views · 9 likes · 1 reposts · 1 replies 01 Sep 2026 · Open on X →

CoT monitoring assumes reasoning traces record what shapes an answer. Yet, we found that models are less likely to verbalize a cue in its CoT when the cue comes from a tool return instead of the user message! 👀 🧵 https://t.co/YXGv7hJVC2 268.9K views · 119 likes · 13 reposts · 3 replies 01 Sep 2026 · Open on X →