MemEvoBench is presented as the first standardized benchmark for long-horizon memory safety in LLM agents, covering adversarial memory injection, noisy tool outputs, and biased feedback across QA and workflow tasks.
hub Canonical reference
Qwen3 technical report
Canonical reference. 75% of citing Pith papers cite this work as background.
30
Pith papers citing it
Background
75% of classified citations
hub tools
citation-role summary
background 6
baseline 1
method 1
citation-polarity summary
representative citing papers
LLM agents often fail to abstain at the right time in uncertain multi-turn tasks, and the CONVOLVE context engineering method raises timely abstention rates on WebShop from 26.7 to 57.4 without parameter updates.
CosFlyTrack provides 12,000 expert UAV trajectories with aligned RGB, depth, segmentation, pose, target state, and bilingual instructions to train visual tracking agents, yielding 53-69 point gains in success rate after fine-tuning.
A unified measurement pipeline on 27 LLM checkpoints shows activation maxima spanning four orders of magnitude, with MoE models 14-23x lower than matched dense models and residual streams carrying the global max in most cases.
Dynamic Latent Routing jointly learns discrete latent codes, routing policies, and model parameters via dynamic search to match or exceed supervised fine-tuning by 6.6 points on average in low-data settings across four datasets and six models.
PPol uses LLM-driven evolutionary program search to create diverse human-like user personas for simulators, yielding 33-62% fitness gains and +17% agent task success on retail and airline domains.
WebEye benchmark and Pixel-Searcher agent enable visual perception tasks by using web search to resolve object identities before precise localization or answering.
TIE derives a sinc-based interval encoding from Temporal Integrability and Duration Invariance principles, raising human-verified temporal constraint satisfaction from 77.34% to 96.03% while preserving visual quality in DiT models.
EquiMem calibrates shared memory in multi-agent debate by computing a game-theoretic equilibrium from agent queries and paths, outperforming heuristics and LLM validators across benchmarks while remaining robust to adversarial agents.
SAM 3D Animal is the first promptable framework for multi-animal 3D reconstruction from single images, built on SMAL+ and trained on the new Herd3D dataset, achieving SOTA results on Animal3D, APTv2, and Animal Kingdom benchmarks.
An automated fact-check-based pipeline for in-the-wild AI image data, when mixed with generator data in continual learning, lets detectors adapt to new generators while avoiding forgetting and delivers 8-9% accuracy gains on two existing models.
Introduces an interactive episodic memory task with user feedback and a Feedback Alignment Module that improves retrieval accuracy on video benchmarks while remaining efficient.
LLM deobfuscation of binaries to pseudocode depends more on reasoning ability and task-specific fine-tuning than on model size, with reasoning models showing robustness across ISAs and obfuscation levels on the new BinDeObfBench.
GraphSSR introduces an adaptive SSR pipeline with SSR-SFT data synthesis and SSR-RL (Authenticity-Reinforced and Denoising-Reinforced stages) to overcome one-size-fits-all subgraph noise in zero-shot LLM graph reasoning.
Dynamic tool synthesis for agentic VideoQA improves accuracy and cuts visible tool calls, but shifts rather than reduces total inference cost, as revealed by a six-group paired audit protocol.
Under matched lost-correct risk, learned early exits beat calibrated scalar exits on free-form math, lose on multiple-choice, and neither certifies savings on small hard sets; serving regime can reverse the savings.
ReactiveGWM introduces a decoupled diffusion architecture for player-NPC interactions that learns game-agnostic response logic for zero-shot strategy transfer across games.
InsightReplay improves long CoT reasoning by extracting critical insights from the trace and replaying them near the active frontier, delivering +1.65 average accuracy gain across 24 model-benchmark settings.
Federated PEFT on LLMs across healthcare and finance datasets performs close to centralized training and beats isolated local training under non-IID conditions.
GRIEF fuzzer finds 15 vulnerabilities including 2 CVEs in vLLM and SGLang by testing concurrent workloads for KV-cache isolation failures and cross-request interference.
Hint Tuning reduces token usage 24-66% (31.5% avg) in reasoning models via 1K self-annotated samples aligned to an instruct model's capabilities while keeping benchmark accuracy.
ξ-DPO rewrites the preference objective as minimizing distance to optimal margins and defines reward as a chosen-to-rejected ratio, yielding a bounded, interpretable margin ξ set directly from the initial reward-gap distribution.
LLM agents trained with a task-success reward on self-generated knowledge can spontaneously explore and adapt to new environments without any rewards or instructions at inference, yielding 20% gains on web tasks and allowing a 14B model to beat Gemini-2.5-Flash.
Prometheus reverse-engineers BDD-style executable specifications from bug failures and uses an RQA validation loop to achieve 93.97% correct patch rate on 680 Defects4J defects while rescuing 74.4% of bugs missed by strong baseline agents.
citing papers explorer
-
MemEvoBench: Benchmarking Safety Risks from Memory Misevolution in LLM Agents
MemEvoBench is presented as the first standardized benchmark for long-horizon memory safety in LLM agents, covering adversarial memory injection, noisy tool outputs, and biased feedback across QA and workflow tasks.
-
Agentic Abstention: Do Agents Know When to Stop Instead of Act?
LLM agents often fail to abstain at the right time in uncertain multi-turn tasks, and the CONVOLVE context engineering method raises timely abstention rates on WebShop from 26.7 to 57.4 without parameter updates.
-
CosFly-Track: A Large-Scale Multi-Modal Dataset for UAV Visual Tracking via Multi-Constraint Trajectory Optimization
CosFlyTrack provides 12,000 expert UAV trajectories with aligned RGB, depth, segmentation, pose, target state, and bilingual instructions to train visual tracking agents, yielding 53-69 point gains in success rate after fine-tuning.
-
Measuring Maximum Activations in Open Large Language Models
A unified measurement pipeline on 27 LLM checkpoints shows activation maxima spanning four orders of magnitude, with MoE models 14-23x lower than matched dense models and residual streams carrying the global max in most cases.
-
Dynamic Latent Routing
Dynamic Latent Routing jointly learns discrete latent codes, routing policies, and model parameters via dynamic search to match or exceed supervised fine-tuning by 6.6 points on average in low-data settings across four datasets and six models.
-
Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents
PPol uses LLM-driven evolutionary program search to create diverse human-like user personas for simulators, yielding 33-62% fitness gains and +17% agent task success on retail and airline domains.
-
From Web to Pixels: Bringing Agentic Search into Visual Perception
WebEye benchmark and Pixel-Searcher agent enable visual perception tasks by using web search to resolve object identities before precise localization or answering.
-
TIE: Time Interval Encoding for Video Generation over Events
TIE derives a sinc-based interval encoding from Temporal Integrability and Duration Invariance principles, raising human-verified temporal constraint satisfaction from 77.34% to 96.03% while preserving visual quality in DiT models.
-
EquiMem: Calibrating Shared Memory in Multi-Agent Debate via Game-Theoretic Equilibrium
EquiMem calibrates shared memory in multi-agent debate by computing a game-theoretic equilibrium from agent queries and paths, outperforming heuristics and LLM validators across benchmarks while remaining robust to adversarial agents.
-
SAM 3D Animal: Promptable Animal 3D Reconstruction from Images in the Wild
SAM 3D Animal is the first promptable framework for multi-animal 3D reconstruction from single images, built on SMAL+ and trained on the new Herd3D dataset, achieving SOTA results on Animal3D, APTv2, and Animal Kingdom benchmarks.
-
Automated In-the-Wild Data Collection for Continual AI Generated Image Detection
An automated fact-check-based pipeline for in-the-wild AI image data, when mixed with generator data in continual learning, lets detectors adapt to new generators while avoiding forgetting and delivers 8-9% accuracy gains on two existing models.
-
Interactive Episodic Memory with User Feedback
Introduces an interactive episodic memory task with user feedback and a Feedback Alignment Module that improves retrieval accuracy on video benchmarks while remaining efficient.
-
Can LLMs Deobfuscate Binary Code? A Systematic Analysis of Large Language Models into Pseudocode Deobfuscation
LLM deobfuscation of binaries to pseudocode depends more on reasoning ability and task-specific fine-tuning than on model size, with reasoning models showing robustness across ISAs and obfuscation levels on the new BinDeObfBench.
-
Beyond One-Size-Fits-All: Adaptive Subgraph Denoising for Zero-Shot Graph Learning with Large Language Models
GraphSSR introduces an adaptive SSR pipeline with SSR-SFT data synthesis and SSR-RL (Authenticity-Reinforced and Denoising-Reinforced stages) to overcome one-size-fits-all subgraph noise in zero-shot LLM graph reasoning.
-
A Cost-Aware, Paired Protocol for Auditing Dynamic Tool Synthesis in Agentic Video Question Answering
Dynamic tool synthesis for agentic VideoQA improves accuracy and cuts visible tool calls, but shifts rather than reduces total inference cost, as revealed by a six-group paired audit protocol.
-
When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models
Under matched lost-correct risk, learned early exits beat calibrated scalar exits on free-form math, lose on multiple-choice, and neither certifies savings on small hard sets; serving regime can reverse the savings.
-
ReactiveGWM: Steering NPC in Reactive Game World Models
ReactiveGWM introduces a decoupled diffusion architecture for player-NPC interactions that learns game-agnostic response logic for zero-shot strategy transfer across games.
-
Stateful Reasoning via Insight Replay
InsightReplay improves long CoT reasoning by extracting critical insights from the trace and replaying them near the active frontier, delivering +1.65 average accuracy gain across 24 model-benchmark settings.
-
Towards the Next Frontier of LLMs, Training on Private Data: A Cross-Domain Benchmark for Federated Fine-Tuning
Federated PEFT on LLMs across healthcare and finance datasets performs close to centralized training and beats isolated local training under non-IID conditions.
-
Continuous Discovery of Vulnerabilities in LLM Serving Systems with Fuzzing
GRIEF fuzzer finds 15 vulnerabilities including 2 CVEs in vLLM and SGLang by testing concurrent workloads for KV-cache isolation failures and cross-request interference.
-
Hint Tuning: Less Data Makes Better Reasoners
Hint Tuning reduces token usage 24-66% (31.5% avg) in reasoning models via 1K self-annotated samples aligned to an instruct model's capabilities while keeping benchmark accuracy.
-
$\xi$-DPO: Direct Preference Optimization via Ratio Reward Margin
ξ-DPO rewrites the preference objective as minimizing distance to optimal margins and defines reward as a chosen-to-rejected ratio, yielding a bounded, interpretable margin ξ set directly from the initial reward-gap distribution.
-
Training LLM Agents for Spontaneous, Reward-Free Self-Evolution via World Knowledge Exploration
LLM agents trained with a task-success reward on self-generated knowledge can spontaneously explore and adapt to new environments without any rewards or instructions at inference, yielding 20% gains on web tasks and allowing a 14B model to beat Gemini-2.5-Flash.
-
Project Prometheus: Bridging the Intent Gap in Agentic Program Repair via Reverse-Engineered Executable Specifications
Prometheus reverse-engineers BDD-style executable specifications from bug failures and uses an RQA validation loop to achieve 93.97% correct patch rate on 680 Defects4J defects while rescuing 74.4% of bugs missed by strong baseline agents.
-
ForestPrune: High-ratio Visual Token Compression for Video Multimodal Large Language Models via Spatial-Temporal Forest Modeling
ForestPrune prunes 90% of visual tokens in video MLLMs like LLaVA-OneVision while retaining 95.8% accuracy by modeling tokens as spatial-temporal forests and scoring importance via tree depth and node roles.
-
MiMuon: Mixed Muon Optimizer with Improved Generalization for Large Models
MiMuon is a hybrid optimizer that achieves a generalization error bound of O(1/N) independent of the small singular-value gap that limits the original Muon bound, while retaining the same O(1/T^{1/4}) convergence rate.
-
LERA: LLM-Enhanced RAG for Ad Auction in Generative Chatbots
LERA is a retrieve-then-generate auction system that refines ad candidate ranking with LLM logits and applies a threshold-aware critical-value payment rule to maintain truthfulness in chatbot ad insertion.
-
Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution
Squeeze Evolve is a multi-model orchestration framework that improves efficiency and performance in verifier-free evolutionary inference, cutting costs up to 3x while matching verifier-based methods on several benchmarks.
-
CoLLM-NAS: Collaborative Large Language Models for Efficient Knowledge-Guided Neural Architecture Search
CoLLM-NAS introduces a collaborative two-LLM framework with Navigator, Generator, and Coordinator modules to perform knowledge-guided neural architecture search, reporting state-of-the-art results on ImageNet and NAS-Bench-201 with 4-10x lower search cost.
-
Retrieval-Augmented Code Generation: A Survey with Focus on Repository-Level Approaches
The paper organizes repository-level retrieval-augmented code generation into a unified framework covering retrieval substrate, control regime, and evaluation setting while summarizing strategies, datasets, and challenges.