DISA decouples partition function estimation using offline importance sampling for distribution-matching LLM-RL, matching or exceeding online baselines like FlowRL on math and code benchmarks while retaining more strategy diversity.
super hub Canonical reference
Advances in neural information processing systems , volume=
Canonical reference. 88% of citing Pith papers cite this work as background.
106
Pith papers citing it
Background
88% of classified citations
hub tools
citation-role summary
background 8
citation-polarity summary
claims ledger
- background An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. Junwei Zhang, Zhongxin Liu, Xing Hu, Xin Xia, and Shanping Li. Vulnerability detection by learning from syntax-based execution paths of code.IEEE Transactions on Software Engineering, 49(8): 4196-4212, 2023. Yaqin Zhou, Shangqing Liu, Jingkai Siow, Xiaoning Du, and Yang Liu. Devign: Effective vulner- ability id
- background Answer:from typing import Listdef median(l: List[int]) -> float: if not l:raise ValueError("The list is empty.")l.sort()n = len(l)mid = n / / 2if n % 2 == 0:return (l[mid -1] + l[mid]) / 2.0else:return float(l[mid]) Queryfrom typing import Listdef median(l: List[int]) -> float:"""Return median of elements in the list l.>>> median([3, 1, 2, 4, 5])3>>> median([-10, 4, 6, 1000, 10, 20])15.0""" Algorithm Designer Test Analyst Algorithm Designer (f)Sampled case in HumanEval. Figure 7.Case study of th
- background distinct trajectories and prevents premature path collapse. As paths diverge, inter-path interaction is gradually attenuated and eventually halted, al- lowing coherent reasoning trajectories to evolve without forced separation. To evaluate the reliability of each generated tra- jectory, we compute its perplexity based on the sequence probability: ppl(y) = exp − 1 L LX t=1 logP(y t |y <t, q) ! (7) where L denotes the trajectory length. During de- coding, paths whose perplexity exceeds a threshold
- background (pscs +nD L)(9) To analyze a concrete scenario, let's assume we can choose an sLM such that its capability is a fraction of the LLM's,i.e., ps = pL n . Using the scaling law from Assumption 3 (pM =αc β M), we can relate the costs: cs = ps α 1/β = pL nα 1/β = n−1/βcL. Substituting these into the heterogeneous cost equation 9: E[CostHeterogeneous](10) = pLcs 2 + pL −p s pL (pscs +nD L)(11) = pLn−1/βcL 2 + n−1 n pL n n−1/βcL +nD L (12) = 1 2 + n−1 n2 n−1/βpLcL + (n−1)D L (13) For the heter
- background TIR, we conduct experiments on domains beyond mathematics. Specifically, we evaluate PRUNETIR on the GPQA-diamond dataset. GPQA-diamond is the highest-quality subset of GPQA (Rein et al., 13 A Case from AIME24 Illustrating Degraded Reasoning in LLMs Problem: Define $f(x)=|| x|-\\tfrac{1}{2}|$ and $g(x)=|| x|-\\tfrac{1}{4}|$. Find the number of intersections of the graphs of \\[y=4 g(f(\\sin (2 \\pi x))) \\quad\\text{ and }\\quad x=4 g(f(\\cos (3 \\pi y))).\\] Solution: Okay, let's try to solve t
- background i . Advantages ˆAdistill i are normalized separately from those of utilization since the two rewards measure different aspects of same outcomes: J distill(θ) =J GRPO θ;{s new,1, . . . , snew,G},{ ˆAdistill 1 , . . . , ˆAdistill G } .(10) Total objective.All terms are combined in a single update: J(θ) =J util(θ) +λ 1 J rerank(θ) +λ 2 J distill(θ).(11) The utility score U(s) is updated non-parametrically via Eq. (5). The full procedure is summarized in Algorithm 1. Training hyperparameter settin
authors
co-cited works
roles
background 8representative citing papers
BOOKMARKS introduces searchable bookmarks as reusable answers to storyline questions, enabling active initialization and passive synchronization for more consistent role-playing agent memory than recurrent summarization.
AWARE augments generative next-POI recommendation with LLM agents that produce user-anchored narratives capturing events, culture, and trends, delivering up to 12.4% relative gains on three real datasets.
OLIVIA treats LLM agent action selection as a contextual linear bandit over frozen hidden states and applies UCB exploration to adapt online, yielding consistent gains over static ReAct and prompt-based baselines on four benchmarks.
LLMs copy biased analyst ratings in investment decisions but a new detection method encourages independent reasoning and can improve stock return predictions beyond human levels.
CMIB uses a conditional multimodal information bottleneck to create reusable agent skills that separate verbalizable text content from predictive perceptual residuals, improving execution stability.
POPO uses bounded importance sampling on positive rollouts and a siamese policy network to achieve implicit negative gradients and stable optimization, matching or exceeding GRPO on math benchmarks such as 36.67% on AIME 2025.
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
LOVER creates an unsupervised logic-regularized verifier that reaches 95% of supervised verifier performance on reasoning tasks across 10 datasets.
NoisyCausal benchmark tests LLMs on causal reasoning with structured noise, and a modular LLM-plus-causal-graph framework outperforms baselines while generalizing to Cladder.
VAnim creates open-domain text-to-SVG animations via sparse state updates on a persistent DOM tree, identification-first planning, and rendering-aware RL with a new 134k-example benchmark.
RSAT uses SFT on verified traces followed by GRPO with NLI faithfulness rewards to make 1-8B models produce verifiable table reasoning with cell citations, raising faithfulness 3.7x to 0.826.
OptiVerse is a new benchmark spanning neglected optimization domains that shows LLMs suffer sharp accuracy drops on hard problems due to modeling and logic errors, with a Dual-View Auditor Agent proposed to improve performance.
LLM tutors leak answers under adversarial student attacks, but a fine-tuned jailbreak agent and simple defenses can benchmark and improve robustness.
A method using predicted rectification difficulty for optimal human sample allocation in LLM-augmented surveys captures 61-79% of theoretical efficiency gains and reduces MSE by 11% on two datasets without pilot data.
GaLa uses hypergraph representations of objects and a TriView encoder with contrastive learning to improve vision-language models on procedural planning benchmarks.
Validity-calibrated reasoning distillation improves transfer of reasoning skills by modulating updates based on relative local validity of next steps instead of enforcing full trajectory imitation.
GraphBit is a DAG-based engine-orchestrated framework for agentic LLMs that achieves 67.6% accuracy with zero hallucinations on GAIA benchmarks.
o1-like models overthink easy tasks; self-training reduces compute use without accuracy loss on GSM8K, MATH500, GPQA, and AIME.
ASB is a new benchmark that tests 10 prompt injection attacks, memory poisoning, a novel Plan-of-Thought backdoor attack, and 11 defenses on LLM agents across 13 models, finding attack success rates up to 84.3% and limited defense effectiveness.
Meta Agent Search uses a meta-agent to iteratively program novel agentic systems in code, producing agents that outperform state-of-the-art hand-designed ones across coding, science, and math while transferring across domains and models.
Activation Addition steers language models by adding contrastive activation vectors from prompt pairs to control high-level properties like sentiment and toxicity at inference time without training.
COALA applies convex optimization reformulations of neural networks to direct preference optimization, claiming single-GPU training with ~18% of DPO's TFLOPs and competitive performance on multiple datasets and models up to 8B parameters.
PACE coordinates low-risk prompt evolution with validated higher-risk control-logic updates to improve frozen SLM agents on benchmarks without model retraining.
citing papers explorer
-
DISA: Offline Importance Sampling for Distribution-Matching LLM-RL
DISA decouples partition function estimation using offline importance sampling for distribution-matching LLM-RL, matching or exceeding online baselines like FlowRL on math and code benchmarks while retaining more strategy diversity.
-
Skill-CMIB: Multimodal Agent Skill for Consistent Action via Conditional Multimodal Information Bottleneck
CMIB uses a conditional multimodal information bottleneck to create reusable agent skills that separate verbalizable text content from predictive perceptual residuals, improving execution stability.
-
Validity-Calibrated Reasoning Distillation
Validity-calibrated reasoning distillation improves transfer of reasoning skills by modulating updates based on relative local validity of next steps instead of enforcing full trajectory imitation.
-
Convex Optimization for Alignment and Preference Learning on a Single GPU
COALA applies convex optimization reformulations of neural networks to direct preference optimization, claiming single-GPU training with ~18% of DPO's TFLOPs and competitive performance on multiple datasets and models up to 8B parameters.
-
PACE: Two-Timescale Self-Evolution for Small Language Model Agents
PACE coordinates low-risk prompt evolution with validated higher-risk control-logic updates to improve frozen SLM agents on benchmarks without model retraining.
-
Distributional Energy-Based Models for Uncertainty-Aware Structured LLM Reasoning
A 149M-parameter distributional energy-based verifier with low-rank adapter ensemble reduces constraint violations in structured LLM reasoning and outperforms or matches much larger models on five benchmarks.
-
Interpretability Can Be Actionable
Interpretability research should be judged by actionability—the degree to which its insights support concrete decisions and interventions—rather than explanatory power alone.
-
Relative Kinetic Utility for Reasoning-Aware Structural Pruning in Large Language Models
RKU is a curvature-aware structural pruning framework that improves LLM reasoning accuracy at 40% sparsity, reaching 13.34% on GSM8K while outperforming baselines and better preserving out-of-distribution representations.
-
TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination
TeamTR is a trust-region framework for multi-agent LLM fine-tuning that resamples trajectories after each update to convert quadratic compounding occupancy shift into linear scaling and yields per-update improvement lower bounds.
-
CAP: Controllable Alignment Prompting for Unlearning in LLMs
CAP is a reinforcement-learning-driven prompt optimization framework that suppresses target knowledge in LLMs while preserving general capabilities, enabling reversible unlearning without any parameter updates.
-
Muon is Scalable for LLM Training
Muon optimizer with weight decay and update scaling achieves ~2x efficiency over AdamW for large LLMs, shown via the Moonlight 3B/16B MoE model trained on 5.7T tokens.
-
The Illusion of Reasoning: Exposing Evasive Data Contamination in LLMs via Zero-CoT Truncation
ZCP detects direct and evasive data contamination in LLMs by truncating CoT reasoning and contrasting zero-CoT accuracy on original versus perturbed isomorphic datasets, plus a Contamination Confidence metric.
-
D$^2$Evo: Dual Difficulty-Aware Self-Evolution for Data-Efficient Reinforcement Learning
D²Evo mines medium-difficulty anchors from the current model, trains a Questioner to generate matching questions, and jointly optimizes Solver and Questioner for progressive gains, outperforming baselines on math reasoning with under 2K real samples.
-
Measure Twice, Click Once: Co-evolving Proposer and Visual Critic via Reinforcement Learning for GUI Grounding
A co-evolving proposer-critic RL framework improves GUI grounding accuracy by letting the model critique its own proposals rendered on screenshots.
-
FedProxy: Federated Fine-Tuning of LLMs via Proxy SLMs and Heterogeneity-Aware Fusion
FedProxy replaces weak adapters with a proxy SLM for federated LLM fine-tuning, outperforming prior methods and approaching centralized performance via compression, heterogeneity-aware aggregation, and training-free fusion.
-
LEPO: Latent Reasoning Policy Optimization for Large Language Models
LEPO applies RL to continuous latent representations in LLMs by injecting Gumbel-Softmax stochasticity for diverse trajectory sampling and unified gradient estimation, outperforming existing discrete and latent RL methods.
-
CHESS: Contextual Harnessing for Efficient SQL Synthesis
CHESS deploys four LLM agents to retrieve information, prune schemas, generate refined SQL candidates, and validate via unit tests, reporting up to 71.10% accuracy on BIRD with 83% fewer calls than leading proprietary baselines.
-
Exploring Pass-Rate Reward in Reinforcement Learning for Code Generation
Pass-rate rewards in critic-free RL for code generation fail to outperform binary rewards because partial-pass solutions induce conflicting gradient directions that do not consistently favor full correctness.
-
Supplement Generation Training for Enhancing Agentic Task Performance
SGT trains a lightweight model to generate task-specific supplemental text that improves performance of a larger frozen LLM on agentic tasks without modifying the large model.
- REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak
- Differentiable Mixture-of-Agents Incentivizes Swarm Intelligence of Large Language Models
- InvEvolve: Evolving White-Box Inventory Policies via Large Language Models with Performance Guarantees