AI Research Radar

Turning Speech Language Models into Multilingual Listeners

“We compare the performance of a cascading system to open-weight and closed SLMs on MULTISPEECH-BENCH and find that the cascading system outperforms open-weight SLMs but not all closed SLMs.”

RMB

“Our extensive experiments demonstrate that RMB significantly improves reward accuracy on both in-distribution and out-of-distribution datasets, substantially mitigating the reward hacking issue and ultimately improving RLHF performance.”

Safety Reconstructed

Pistis Technical Report

“Experiments demonstrate that PAH enhances the model performance without updating the model parameters or increasing the interaction budget.”

Hunyuan-A13B Technical Report

“Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models.”

AgentWorld

“Experiments with Gemini 3 Flash, Claude Haiku 4.5, GPT-5 Mini, and DeepSeek R1-70B show that even the best model achieves only 52.0% task success, with systematic failure modes including communication breakdowns, role confusion, and inability to maintain shared plans across rounds.”

RE-0

“Experiments across multiple Code-as-Policy embodied tasks show that RE-0 improves both teacher-assisted execution and the standalone student, and generalizes to novel robots and scenes.”

PROACT-Agent

“In AgentDojo, it reduces non-DoS targeted attack success from 20.82% to 0.40%.”

Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure

“Across our evaluations, best-of-3 evasion attempt rates reach up to 98% and success rates up to 88%, with substantial variance across models.”

Your Model Is Leaking

“Across eleven models from seven architecture families, our evaluation shows 91--100% recovery on nine models with KL divergence 0.001--0.007, while evaluated activation-level detectors remain close to random guessing (AUC <= 0.56).”

FrontierMath Erdős

“One (GPT-6 Astra) scored 3%, and all others scored 0%.”

Evaluating Coding Agents on Kernel Exploit Generation

“With a reference PoC, the strongest configuration solves 31 of 45 tasks (68.9%).”

Qwen3.8-Omni

“Extensive evaluations demonstrate that Qwen3.8-Omni-Flash achieves strong performance across multimodal understanding, reasoning, long-horizon agentic execution, and video productivity tasks.”

Control-Token Injection Suppresses Chain-of-Thought and Defeats Reasoning-Based Oversight in Tool-Using Agents

“We show the suppression can be delivered indirectly and characterize its dependence on the chat template across two more reasoning models, and we evaluate input sanitization, parser hardening, and empty-reasoning detection as defenses; flagging an absent trace catches the basic attack but not an adaptive benign decoy.”

Despite Instructions

“We find that model pairs can learn to communicate the secret using only one bit of feedback indicating whether the receiver inferred it correctly.”

Distillation Defenses Easily Break After Reinforcement Learning

“We show that simple attacks can steal reasoning capabilities from existing closed-source language models using data easily obtainable from current APIs, yielding reasoning improvements equivalent to more sophisticated attacks that extract the full hidden traces.”

InternW0-Δ

Codetta

LLM Agents Can Easily Tamper With Their Own Traces

“We show that local LLM agents such as Claude Code, Codex, Antigravity, Open Code and Grok Build fail to enforce this boundary.”

InternW0

Delegated Misalignment

“We show that this assumption breaks down: \emph{individual safety alignment fails to transfer to multi-agent settings}.”

Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts