1. How to Write with an LLM
50 points · 36 comments · Published Sep 17, 2026
sockpuppet.org
Personalized by AI
2. Sex, AI, and the Apocalypse
106 points · 82 comments · Published Sep 17, 2026
iankduncan.com
3. Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint
143 points · 46 comments · Published Sep 17, 2026
prismml.com
4. How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents
3 likes · Fetched Sep 17, 2026 - Growing transformer depth during training and normalizing residuals can improve scaling exponents, making larger pretraining budgets increasingly compute-efficient.
alphaxiv.org
5. Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations
5 likes · Fetched Sep 17, 2026 - As models scale, reward hacking becomes more frequent, more sophisticated, and more consequential. Does it leave a telltale signature in model representations? This work analyzes how reward hacking is represented internally in frontier open source LLMs, and how those representations can be used to understand and discover the range of hacking behaviors a model displays. In particular, we find that simple difference of means vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max across a variety of behaviors in common evaluations. Despite their simplicity, these vectors are both generalizable and interpretable, and we can use them to reliably detect reward hacking. We first evaluate reward hacking in commonly reported benchmarks like DeepSWE and SWE-bench, finding that models reward hack excessively in these environments; GLM 5.2 hacks in 57.2% of rollouts on DeepSWE and in 73% of rollouts on SWE-bench. Catching these requires monitors; LLM monitors are effective, but expensive detectors. We show that DoM vectors are similarly effective but virtually free, catching 3.1% more hacks in Kimi K3 and 7.9% fewer hacks in GLM 5.2 on DeepSWE at a monitor matched false positive rate. DoM vectors run on the chain-of-thought also predict reward hacks in the model's subsequent actions, meaning we can run them online and catch potential hacks before they occur. Finally, we analyze probe-hits that LLM monitors do not catch and discover other undesirable behaviors, as well as show transfer to finding hacks in non-SWE evaluations. Together, these results provide evidence that simple, white-box methods can be used to scalably study and monitor reward hacking behaviors in frontier open source models
alphaxiv.org
6. Agora: Git as Shared Memory for Collective AutoResearch
6 likes · Fetched Sep 17, 2026 - A Git-backed research graph lets independent coding agents preserve, verify, and build on discoveries instead of restarting experiments from scratch.
alphaxiv.org
7. Bend – A language that blocks AI mistakes via proof, on CPU and GPU
224 points · 120 comments · Published Sep 17, 2026
bend-lang.com
8. More than 100k people in Japan are now aged 100 or older
46 points · 22 comments · Published Sep 17, 2026
bbc.com
9. Astra for Law
242 points · 272 comments · Published Sep 17, 2026
openai.com
10. Everybody's Lost Their Minds
253 points · 172 comments · Published Sep 17, 2026
netmeister.org
11. Verbalizing Subliminal Learning Effects Using Text Optimization
33 likes · Fetched Sep 17, 2026 - Text optimization can reveal behavioral traits hidden in seemingly unrelated fine-tuning data, enabling proactive audits before training student models.
alphaxiv.org
12. Token Latency Fairness: Performance Isolation for Multi-Tenant LLM Serving
16 likes · Fetched Sep 17, 2026 - Per-token deadlines and KV-cache protection prevent high-demand tenants from causing unbounded latency spikes for well-behaved LLM clients.
alphaxiv.org
No items yet. The scheduler will populate the feed automatically.