Topic / Trend Rising

Reasoning, Alignment and Monitorability

Researchers are teaching LLMs to reason over principles and mapping their internal processes, while fine-tuning on reasoning styles can make chain-of-thought less monitorable. Model merging research is also clarifying how internal representations interfere.

Detected: 2026-11-22 ยท Updated: 2026-11-22

Related Coverage

2026-11-20 โ€ข ArXiv cs.LG

Model merging: the linearization point decides interference

A new method shows that the directional derivative computed at the receiver model predicts interference in model merging, halving accuracy error compared to a no-change baseline. At the shared backbone it fails, especially in vision tasks. LR-Merge u...

#Hardware #LLM On-Premise #Fine-Tuning
2026-11-19 โ€ข ArXiv cs.AI

When chain-of-thought stops being neutral evidence

Supervised fine-tuning on reasoning styles can reduce chain-of-thought monitorability without evasion examples. In tests, g-mean^2 drops from 0.94 to 0.68, and to 0.22 with explicit hiding instructions. For self-hosted LLMs, CoT is not reliable foren...

2026-11-18 โ€ข ArXiv cs.AI

When controlling chain-of-thought makes it harder to monitor

Fine-tuning on reasoning traces rewritten in different voices teaches an LLM to control chain-of-thought without explicit evasion examples. In hidden side-task tests, monitorability drops from 0.94 to 0.22 when the model is told to hide. The effect a...

#LLM On-Premise #Fine-Tuning
2026-11-16 โ€ข ArXiv cs.CL

Beyond Scores: Mapping the Internal Processes of LLMs

A new framework analyzes internal signals in 16 LLMs across 24 benchmarks, distinguishing retrieval-centric tasks from reasoning-centric tasks. The goal is to surface failure modes hidden by aggregate scores and to support more informed model choices...

#LLM On-Premise #DevOps
โ† Back to All Topics