Latest AI News & Industry Updates

Debiasing LLM-as-a-Judge evaluation with TraceJudgeBench

This paper provides a detailed analysis of how debiasing methods affect the accuracy of LLMs used as judges, which is important for anyone working with RAG or agent-workflow evaluation.

#LLM-as-a-judge #debiasing #RAG #evaluation

arXiv: DeepSeek@Liang Zhao et al. (3)Source

Papers US vs ChinaOfficial10:44

Chinese content上海 AGI Bar 的 Next Token 第二期上线，讨论 GPT-6 Astra、DeepSeek V4.1 Flash 和苹果的 iPhone Duo

朋友录的播客，周一摸鱼时可以听，内容有 GPT-6、DeepSeek V4.1 和苹果的 iPhone Duo 这些热点。

#GPT-6 #DeepSeek #iPhone Duo #AGI Bar

Event

官方一手歸藏(guizang.ai)@op741811 sourcesSource

Papers BuildersMulti-sourceSelected78°11:02

PARSER agent design for long-context reasoning

PARSER's design could lead to more efficient and accurate AI assistants for processing long documents, which is relevant for any user dealing with large amounts of text.

#PARSER #long-context #reasoning models #multi-hop QA

Event

elvis@omarsar03 sourcesSource

Papers BuildersMulti-sourceSelected09:11

AutoTrans Framework for RISC-V Security Assertions

AutoTrans reduces hardware security verification time from hours to minutes per assertion, making it valuable for RISC-V developers in ASEAN.

#AutoTrans #RISC-V #Deepseek V4 #hardware security

Event

arXiv: DeepSeek@Sharjeel Imtiaz et al. (3)3 sourcesSource

Papers BuildersMulti-source10:33

Chinese contentRAME框架在谷物育种信息提取任务中夺冠

DeepSeek团队用RAME框架在谷物育种信息提取任务击败GPT-5.5，代码已开源。

#RAME #DeepSeek-V4-Flash #GPT-5.5 #信息提取

Event

arXiv: DeepSeek@Hang Zhao, Jiahao Wang4 sourcesSource

Papers BuildersMulti-sourceSelected73°10:32

SRD-GUARD defense framework for LLMs

Developers using LLMs in safety-critical applications can implement SRD-GUARD to detect concealed harmful intent while minimizing false positives compared to existing defenses.

#SRD-GUARD #Llama-3-8B-Uncensored #DeepSeek-V4-Flash #jailbreak defense

Event

arXiv: DeepSeek@Qi Wang et al. (3)3 sourcesSource

Papers BuildersMulti-sourceSelected09:23

DeepSeek-V4-Flash Residual Stream Analysis

The findings show that DeepSeek-V4-Flash uses only part of its four-stream capacity, with early mixing being crucial and late mixing providing minimal benefit.

#DeepSeek-V4-Flash #residual streams #mHC #attention

Event

arXiv: DeepSeek@Pengxiang Zhao et al. (5)4 sourcesSource

PapersMulti-sourceSelected73°11:13

KC-Bench Evaluates LLM Agent Conflicts

KC-Bench provides developers with a reproducible diagnostic tool to improve conflict-aware reasoning in AI agents.

#KC-Bench #LLM agents #DeepSeek-V4-Flash #GLM-5.2

Event

arXiv: DeepSeek@Yaxing Lyu et al. (5)5 sourcesSource

PapersSelected73°11:50

Trace as State for Long-Context Transformers

The method significantly boosts performance on long-context reasoning tasks without changing transformer architecture, benefiting developers working with complex document analysis.

#Trace as State #DeepSeek V4 Pro Preview #GLM-5.2 #long-context reasoning

arXiv: DeepSeek@Xu Zou, Jie TangSource

PapersMulti-sourceSelected73°11:49

LLM-Ontology Fusion for Rare Disease Diagnosis

The fusion model improves diagnostic accuracy while preserving evidence trails, crucial for rare diseases where misdiagnosis can have severe consequences.

#LLM #ontology rankers #rare disease diagnosis #DeepSeek-V4-Flash

Event

arXiv: DeepSeek@Zhaoyang Jiang et al. (8)4 sourcesSource

Papers BuildersSelected10:20

Token Count Fingerprinting Study of LLM APIs

Token count fingerprinting works for identifying shared tokenization stacks but fails as a standalone test for model-family lineage, affecting API service differentiation.

#token-count fingerprinting #LLM APIs #model attribution #Qwen 3.8

arXiv: DeepSeek@Bo ChenSource

PapersMulti-sourceSelected73°18:01

Thousand-Graph Hypothesis for Code Reasoning

This method improves code reasoning accuracy in large repositories without expensive pre-training or maintenance burden.

#DeepSeek-V4-Flash #SWE-bench #code reasoning #repository-level

Event

arXiv: DeepSeek@Fei Ding5 sourcesSource

Papers BuildersMulti-source09:18

Chinese contentEVOMAL: 自进化的编码智能体中的自毒化问题

研究揭示了自进化的编码智能体中的自毒化问题，并提出了counter-prompt防御措施，有效降低了恶意技能的传播。对于关注AI安全和智能体研究的读者来说，这是一篇不容错过的论文。

#智能体 #AI安全 #自毒化 #防御措施