Skip to content

๐Ÿ‘ป Hallucination Detection

๐Ÿ’ฌ EMNLP2026 ยท 6 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿง  NeurIPS2026 (19) ยท ๐ŸŽž๏ธ ECCV2026 (20) ยท ๐Ÿ“ท CVPR2026 (33) ยท ๐Ÿ”ฌ ICLR2026 (40) ยท ๐Ÿ’ฌ ACL2026 (28) ยท ๐Ÿงช ICML2026 (21)

๐Ÿ”ฅ Top topics: LLM ร—2

HaluMem: Evaluating Hallucinations in Memory Systems of Agents

HaluMem establishes the first operation-level hallucination evaluation benchmark for LLM agent memory systems, decoupling traditional black-box QA into memory extraction, updating, and question answering stages across dialogues reaching one million tokens, revealing how upstream extraction omissions cascade into downstream updating and QA failures.

Library Hallucinations in LLM-Generated Code: A Risk Analysis Grounded in Developer Queries

This paper presents the first systematic study of how realistic developer prompt variations (natural language descriptors and controlled user typos/fabrications) trigger library name and member hallucinations in LLM-generated code, revealing alarming vulnerabilities to temporal constraints and typos, and introducing LIBHALLUBENCH for reproducible evaluation.

PVFH-Bench: Benchmarking Premise-Induced Visual Fixation Hallucinations in Vision-Language Models

This paper identifies a critical compounded failure mode termed Premise-Induced Visual Fixation Hallucination (PVFH) along with a diagnostic benchmark PVFH-Bench, revealing that contemporary LVLMs anchor onto visually confusable regions under misleading textual premises and that explicit chain-of-thought reasoning exacerbates rather than mitigates this hallucination trap.

Teaching LLMs to Abstain via Fine-Grained Semantic Confidence Reward

Addressing the ambiguous knowledge boundary caused by coarse-grained global signals in LLM abstention alignment, this paper proposes FISCORE, an RL framework that derives sample-specific semantic confidence rewards from clustering cluster sizes via GRPO, and introduces the monotonically sound F1rel reliability metric.

The Trilemma of Truth in Large Language Models

Addressing three flawed assumptions in conventional binary veracity probing, this paper proposes sAwMIL, a multiclass probing framework combining sparse multiple-instance learning with conformal prediction that reliably isolates true, false, and neither signals with causal steering guarantees across 16 LLMs.

To Answer or to Abstain: Mitigating Search-Agent Hallucinations via Abstention-Aware Reinforcement Learning

Addressing the over-answering hallucination issue where search-augmented LLM agents fabricate answers upon retrieval failure, this paper proposes Abstention-aWAre Reinforcement Learning (AWA-RL), dynamically shaping refusal rewards via query-specific prior capability estimation, an optimistic courage factor, and on-policy excess penalty to achieve a superior capability-reliability trade-off.