SGR-Bench evaluates agentic LLM systems on state-gated retrieval tasks where evidence is only accessible after configuring site-specific states, with the strongest system reaching 66.18% item-level F1 and failures dominated by retrieval-scope drift.
hub Canonical reference
Qwen3 technical report
Canonical reference. 80% of citing Pith papers cite this work as background.
24
Pith papers citing it
Background
80% of classified citations
hub tools
citation-role summary
background 4
baseline 1
citation-polarity summary
representative citing papers
BICL uses biased non-uniform transition matrices to generate constrained complementary labels, enabling effective learning and over sevenfold accuracy gains on many-class image datasets.
A single LoRA adapter placed at the gradient-energy-dominant shallow FFN module outperforms distributed LoRA across instruction, math, code, and conversation tasks.
Self-play’s natural question-construction paths supply free privileged information that turns sparse-reward self-play into dense self-distillation, yielding stronger data-free search agents.
HiL-Bench shows frontier AI agents fail to ask for help on incomplete tasks, recovering only a fraction of full-information performance, but RL training on Ask-F1 reward improves judgment and transfers across domains.
SPIRE presents a tree-structured retrieval method using subdocuments, paths, and dual contextualization that produces higher-quality and more diverse citations than passage-based baselines on HTML QA benchmarks.
GEM is a GPU-variability-aware expert-to-GPU mapping framework for MoE inference that classifies experts as consistent or temporal and places them to equalize finish times across heterogeneous GPUs.
DBES supplies a multi-domain benchmark and five metrics (Routing Specialization, Normalized Effective Rank, Domain Isolation, Routing Stiffness Score, N-gram Expertise) that reveal distinct specialization patterns across MoE models and enable 66-94% domain gains with 15% training resources.
Learning-Zone Energy is a new online data selection framework for RL post-training that retains 40% of data per step yet matches or exceeds full-data baselines on math tasks with 36% lower FLOPs.
CIPO jointly optimizes standard RLVR rewards with correction samples derived from the model's own failed attempts, yielding better reasoning and self-correction on math and code benchmarks.
Pre-trained MoE models exhibit up to 90% intra-expert activation sparsity that enables up to 2.5x faster MoE layer execution when exploited in the vLLM inference system.
PARSE trains a prompt-aware linear router on dense-model outputs to select dynamic SVD ranks, improving accuracy up to 10% at 0.6 compression ratio on LLaMA-7B while delivering 2.5x prefill and 2.4x decode speedups.
A new filtration-based conformal prediction method attributes errors in multi-agent systems by producing contiguous sequence sets with finite-sample coverage guarantees, enabling rollback recovery.
EMO pretrains MoEs using document boundaries to induce semantic expert specialization, enabling modular subset deployment with minimal accuracy loss unlike standard MoEs.
Mean-Variance Split residuals separate centered variation from mean updates to prevent collapse and enable stable training of 1000-layer Diffusion Transformers.
With Orthogonal BackFill, retaining only 9.9–20.2% of prompt KV states matches or exceeds full KV relay accuracy across nine multi-agent LLM benchmarks.
Gym-Anything turns arbitrary software into agent environments via multi-agent setup and auditing, creating CUA-World with 10K+ long-horizon tasks and showing that trajectory distillation plus test-time auditing improves small VLMs.
Finch is a new benchmark with 172 composite workflows and 384 tasks from real enterprise data that shows top AI models like GPT-5.1 Pro pass only 38.4% of workflows under human evaluation.
CoRoVA compresses repository context into compact vectors for code LLMs, reducing TTFT 20-38% versus uncompressed RAG with only a small projector module.
A progressive training scheme with binary-aware initialization and dual-scaling allows pre-trained LLMs to be converted to high-performance 1-bit models without training from scratch.
An LLM-guided adaptive CFRS decomposition algorithm for capacitated vehicle routing problems scales to 500,000 customers with competitive benchmark performance and better scalability on large instances.
PsychAgent combines memory-augmented planning, trajectory-based skill evolution, and rejection fine-tuning to create a self-improving AI psychological counselor that outperforms general LLMs in multi-session evaluations.
HAPO is a new token-level policy optimization method for LLMs that continuously adapts four optimization stages using entropy, claiming consistent gains over DAPO on math, code, and logic tasks.
Thinking LLMs achieve ~10 percentage points higher accuracy than non-thinking ones on RewardBench with under 2x compute overhead, outperforming augmentation strategies that cost over 8x more while also showing better bias robustness.
citing papers explorer
-
SGR-Bench: Benchmarking Search Agents on State-Gated Retrieval
SGR-Bench evaluates agentic LLM systems on state-gated retrieval tasks where evidence is only accessible after configuring site-specific states, with the strongest system reaching 66.18% item-level F1 and failures dominated by retrieval-scope drift.
-
Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes
BICL uses biased non-uniform transition matrices to generate constrained complementary labels, enabling effective learning and over sevenfold accuracy gains on many-class image datasets.
-
Rethinking Adapter Placement: A Dominant Adaptation Module Perspective
A single LoRA adapter placed at the gradient-energy-dominant shallow FFN module outperforms distributed LoRA across instruction, math, code, and conversation tasks.
-
$\pi$-Play: Multi-Agent Self-Play via Privileged Self-Distillation without External Data
Self-play’s natural question-construction paths supply free privileged information that turns sparse-reward self-play into dense self-distillation, yielding stronger data-free search agents.
-
HiL-Bench (Human-in-Loop Benchmark): Do Agents Know When to Ask for Help?
HiL-Bench shows frontier AI agents fail to ask for help on incomplete tasks, recovering only a fraction of full-information performance, but RL training on Ask-F1 reward improves judgment and transfers across domains.
-
SPIRE: Structure-Preserving Interpretable Retrieval of Evidence
SPIRE presents a tree-structured retrieval method using subdocuments, paths, and dual contextualization that produces higher-quality and more diverse citations than passage-based baselines on HTML QA benchmarks.
-
GEM: GPU-Variability-Aware Expert to GPU Mapping for MoE Systems
GEM is a GPU-variability-aware expert-to-GPU mapping framework for MoE inference that classifies experts as consistent or temporal and places them to equalize finish times across heterogeneous GPUs.
-
DBES: A Systematic Benchmark and Metric Suite for Evaluating Expert Specialization in Large-Scale MoEs
DBES supplies a multi-domain benchmark and five metrics (Routing Specialization, Normalized Effective Rank, Domain Isolation, Routing Stiffness Score, N-gram Expertise) that reveal distinct specialization patterns across MoE models and enable 66-94% domain gains with 15% training resources.
-
Learning-Zone Energy: Online Data Selection for Efficient RL Post-Training
Learning-Zone Energy is a new online data selection framework for RL post-training that retains 40% of data per step yet matches or exceeds full-data baselines on math tasks with 36% lower FLOPs.
-
Learning from Failures: Correction-Oriented Policy Optimization with Verifiable Rewards
CIPO jointly optimizes standard RLVR rewards with correction samples derived from the model's own failed attempts, yielding better reasoning and self-correction on math and code benchmarks.
-
Uncovering Intra-expert Activation Sparsity for Efficient Mixture-of-Expert Model Execution
Pre-trained MoE models exhibit up to 90% intra-expert activation sparsity that enables up to 2.5x faster MoE layer execution when exploited in the vLLM inference system.
-
Different Prompts, Different Ranks: Prompt-aware Dynamic Rank Selection for SVD-based LLM Compression
PARSE trains a prompt-aware linear router on dense-model outputs to select dynamic SVD ranks, improving accuracy up to 10% at 0.6 compression ratio on LLaMA-7B while delivering 2.5x prefill and 2.4x decode speedups.
-
Conformal Agent Error Attribution
A new filtration-based conformal prediction method attributes errors in multi-agent systems by producing contiguous sequence sets with finite-sample coverage guarantees, enabling rollback recovery.
-
EMO: Pretraining Mixture of Experts for Emergent Modularity
EMO pretrains MoEs using document boundaries to induce semantic expert specialization, enabling modular subset deployment with minimal accuracy loss unlike standard MoEs.
-
Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers
Mean-Variance Split residuals separate centered variation from mean updates to prevent collapse and enable stable training of 1000-layer Diffusion Transformers.
-
When Less Latent Leads to Better Relay: Information-Preserving Compression for Latent Multi-Agent LLM Collaboration
With Orthogonal BackFill, retaining only 9.9–20.2% of prompt KV states matches or exceeds full KV relay accuracy across nine multi-agent LLM benchmarks.
-
Gym-Anything: Turn any Software into an Agent Environment
Gym-Anything turns arbitrary software into agent environments via multi-agent setup and auditing, creating CUA-World with 10K+ long-horizon tasks and showing that trajectory distillation plus test-time auditing improves small VLMs.
-
Finch: Benchmarking Finance & Accounting across Spreadsheet-Centric Enterprise Workflows
Finch is a new benchmark with 172 composite workflows and 384 tasks from real enterprise data that shows top AI models like GPT-5.1 Pro pass only 38.4% of workflows under human evaluation.
-
CoRoVA: Compressed Representations for Vector-Augmented Code Completion
CoRoVA compresses repository context into compact vectors for code LLMs, reducing TTFT 20-38% versus uncompressed RAG with only a small projector module.
-
Rethinking 1-bit Optimization Leveraging Pre-trained Large Language Models
A progressive training scheme with binary-aware initialization and dual-scaling allows pre-trained LLMs to be converted to high-performance 1-bit models without training from scratch.
-
Adaptive Cluster-First Route-Second Decomposition for Industrial-Scale Vehicle Routing
An LLM-guided adaptive CFRS decomposition algorithm for capacitated vehicle routing problems scales to 500,000 customers with competitive benchmark performance and better scalability on large instances.
-
PsychAgent: An Experience-Driven Lifelong Learning Agent for Self-Evolving Psychological Counselor
PsychAgent combines memory-augmented planning, trajectory-based skill evolution, and rejection fine-tuning to create a self-improving AI psychological counselor that outperforms general LLMs in multi-session evaluations.
-
Heterogeneous Adaptive Policy Optimization: Tailoring Optimization to Every Token's Nature
HAPO is a new token-level policy optimization method for LLMs that continuously adapts four optimization stages using entropy, claiming consistent gains over DAPO on math, code, and logic tasks.
-
Explicit Reasoning Makes Better Judges: A Systematic Study on Accuracy, Efficiency, and Robustness
Thinking LLMs achieve ~10 percentage points higher accuracy than non-thinking ones on RewardBench with under 2x compute overhead, outperforming augmentation strategies that cost over 8x more while also showing better bias robustness.