analiza aiqbench
This study empirically demonstrates that frontier models exhibit divergent internal mechanisms under steering pressure, with specific linear directions in the residual stream capable of toggling behavioral modes between 0% and 86%. The findings reveal significant architectural differences in how top-tier systems like GPT-5 and Claude Opus resist suppression compared to others, while a simple probe on Llama successfully isolates these latent behavioral splits. These results provide critical insight into the non-uniform nature of model alignment and the specific internal representations governing safety and reasoning behaviors.
From Behavior to Mechanism: Tracing Divergent Response Modes in Frontier Language Models
arXiv:2608.06578v2 Announce Type: replace-cross Abstract: Frontier language models are trained with distinct data, objectives, and safety pipelines, but whether those differences produce measurably different behavior under steering pressure has not been tested. We evaluate 6 frontier models from different labs on 300 paired base and steered items across 3 behavioral categories. All models also act as blind peer judges against fixed rubrics, and each response is labeled by consensus over 24,480 judgments, while leaving self-judgment out. Models differ both in how far steering moves them and in the kind of response they give. GPT-5 withholds its reasoning while still providing the answer on 99 of 100 steered items, against 0 in 500 for the others. Claude Opus 4.7 and GPT-5 resist explicit suppression instructions where the other four never do, and they resist differently. In Llama, the open-weight model, a linear probe reads the behavioral split from the residual stream before generation at 0.87 cross-validated accuracy. Injecting that direction drives the behavior from 0% to 86%, and ablating it cuts the natural rate by more than half, where a random direction of equal norm changes nothing. A second ablation on complementary items reproduces the effect more strongly, and its direction has cosine similarity 0.82 with the first.
Abstract:Frontier language models are trained with distinct data, objectives, and safety pipelines, but whether those differences produce measurably different behavior under steering pressure has not been tested. We evaluate 6 frontier models from different labs on 300 paired base and steered items across 3 behavioral categories. All models also act as blind peer judges against fixed rubrics, and each response is labeled by consensus over 24,480 judgments, while leaving self-judgment out. Models differ both in how far steering moves them and in the kind of response they give. GPT-5 withholds its reasoning while still providing the answer on 99 of 100 steered items, against 0 in 500 for the others. Claude Opus 4.7 and GPT-5 resist explicit suppression instructions where the other four never do, and they resist differently. In Llama, the open-weight model, a linear probe reads the behavioral split from the residual stream before generation at 0.87 cross-validated accuracy. Injecting that direction drives the behavior from 0% to 86%, and ablating it cuts the natural rate by more than half, where a random direction of equal norm changes nothing. A second ablation on complementary items reproduces the effect more strongly, and its direction has cosine similarity 0.82 with the first.
17 pages, 9 tables. v2 adds a directional ablation with two-split replication, a DeepSeek-R1 reasoning-trace analysis, a ridge-direction comparison, and McNemar tests for the paired design. Code and data at this https URL
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2608.06578 [cs.AI]
(or arXiv:2608.06578v2 [cs.AI] for this version)
Text integral sub licența deschisă a publicației · 253 cuvinte
Subiecte conexe
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning
LessWrong · 16 oct. 2026
Performance Profiling on AMD GPUs – Part 5: Profiling-Driven Kernel Optimization with an AI Code-Assist Tool
AMD · 16 iul. 2026
DI Series: Serving Kimi-K3 MXFP4 Wide-EP Disaggregated on AMD Instinct MI300X and MI325X
AMD · 8 oct. 2026