Expander SAEs apply left-d-regular expander masks to TopK SAEs, learning only dn decoder parameters instead of mn and tracing a storage-fidelity frontier that reaches 293x compression with 84% retained performance on Qwen2.5-3B.
mega hub Mixed citations
Qwen2.5 Technical Report
Mixed citation behavior. Most common role is background (65%).
1157
Pith papers citing it
15
external citations · Pith
Background
65% of classified citations
abstract
In this report, we introduce Qwen2.5, a comprehensive series of large language models (LLMs) designed to meet diverse needs. Compared to previous iterations, Qwen 2.5 has been significantly improved during both the pre-training and post-training stages. In terms of pre-training, we have scaled the high-quality pre-training datasets from the previous 7 trillion tokens to 18 trillion tokens. This provides a strong foundation for common sense, expert knowledge, and reasoning capabilities. In terms of post-training, we implement intricate supervised finetuning with over 1 million samples, as well as multistage reinforcement learning. Post-training techniques enhance human preference, and notably improve long text generation, structural data analysis, and instruction following. To handle diverse and varied use cases effectively, we present Qwen2.5 LLM series in rich sizes. Open-weight offerings include base and instruction-tuned models, with quantized versions available. In addition, for hosted solutions, the proprietary models currently include two mixture-of-experts (MoE) variants: Qwen2.5-Turbo and Qwen2.5-Plus, both available from Alibaba Cloud Model Studio. Qwen2.5 has demonstrated top-tier performance on a wide range of benchmarks evaluating language understanding, reasoning, mathematics, coding, human preference alignment, etc. Specifically, the open-weight flagship Qwen2.5-72B-Instruct outperforms a number of open and proprietary models and demonstrates competitive performance to the state-of-the-art open-weight model, Llama-3-405B-Instruct, which is around 5 times larger. Qwen2.5-Turbo and Qwen2.5-Plus offer superior cost-effectiveness while performing competitively against GPT-4o-mini and GPT-4o respectively. Additionally, as the foundation, Qwen2.5 models have been instrumental in training specialized models such as Qwen2.5-Math, Qwen2.5-Coder, QwQ, and multimodal models.
hub tools
citation-role summary
background 90
method 21
baseline 13
other 8
dataset 7
citation-polarity summary
claims ledger
- abstract In this report, we introduce Qwen2.5, a comprehensive series of large language models (LLMs) designed to meet diverse needs. Compared to previous iterations, Qwen 2.5 has been significantly improved during both the pre-training and post-training stages. In terms of pre-training, we have scaled the high-quality pre-training datasets from the previous 7 trillion tokens to 18 trillion tokens. This provides a strong foundation for common sense, expert knowledge, and reasoning capabilities. In terms of post-training, we implement intricate supervised finetuning with over 1 million samples, as well
authors
mega hub controls
export citing contexts JSON
export graph JSON
export full bundle JSON
open full Pith review
annotated reader queued
Recognition alignment
counterfactual ablation
co-cited works
representative citing papers
TW-LegalBench evaluates 13 LLMs on over 30,000 Taiwanese legal tasks from exams and judgments, showing top models pass lawyer thresholds but struggle with exact statute citations.
Fine-tuning updates frequently stale activation monitors for language model safety while quantization does not, with degradation predictable and repairable via label-free realignment.
Acceptance Cards is a new four-diagnostic standard for safe fine-tuning defense claims that requires statistical reliability, fresh semantic generalization, mechanism alignment, and cross-task transfer; under this protocol SafeLoRA fails the full-card pass on Gemma-2-2B-it.
FormalRewardBench is the first benchmark for reward models in formal theorem proving, consisting of 250 Lean 4 preference pairs that show frontier LLMs scoring 59.8% while specialized provers score only 24.4%.
Creates the first benchmark dataset integrating papers, slides, videos, and presentations for evaluating AI models on fine-grained multimodal correspondences in science.
ArgBench unifies 33 existing datasets into a standardized benchmark for testing LLMs across 46 argumentation tasks and analyzes the impact of prompting techniques and model factors on performance.
RLCracker is a reinforcement learning attack that erases LLM watermarks at 98.5% success rate with minimal data and generalizes across ten schemes and multiple model sizes.
LLaDA is a scalable diffusion-based language model that matches autoregressive LLMs like LLaMA3 8B on tasks and surpasses GPT-4o on reversal poem completion.
Active lifting of data distributions via latent-sequence sampling and max-rate uncertainty reduction formally derives slow-thinking LLMs and places them on representation and sampler hierarchies that can be climbed.
DecompRL is an RL method that learns modular code decomposition for LLMs, enabling exponential candidate generation via recombination to solve harder coding problems with lower GPU cost.
Conditional Co-Ablation recovers self-repair backup heads in transformers by scoring conditional ablation growth, raising ROC-AUC from 0.33 to 0.91 on the IOI circuit and transferring to induction across models.
The paper proposes Multi-Head Recurrent Memory (MHM) with a select-then-update strategy to improve memory retention in long-context recurrent agents.
Static SFT and RL training for tool-use agents leads to performance drops under open-world distributional shifts across perception, interaction, reasoning and internalization; perturbation-augmented fine-tuning is proposed as mitigation.
Multi-agent LLMs classify USPTO reactions and write verified SMIRKS rules, expanding a reaction taxonomy from 68 to 14,073 classes and matching proprietary classifiers on held-out and out-of-distribution data.
TASA improves task-aware mixed-precision LLM quantization by searching calibration data mixtures via gradient-trace alignment and aggregating perplexity plus reasoning sensitivity signals, enabling 3.5-bit models to match or beat 4-bit baselines with over 20-point gains on GSM8K.
HARC couples harmfulness and refusal directions at prompt and response positions, yielding the best robustness-capability-usability trade-off among major safety methods.
Releases SEFORA corpus of instructor feedback on college writing and UniMatch evaluation showing no LLM configuration exceeds 0.4 F1 in matching instructor priorities.
FlexiSLM is the first spoken language model supporting dynamic and controllable frame rates on speech input and output, outperforming fixed-rate 7B models at high quality and enabling faster inference at lower rates like 6.25 Hz.
Cortex uses an Ontological Corpus Graph to structure web-scale corpora, creating a refined 24.14B-token corpus and a new benchmark validated on eight LLMs.
Fuzzing via Gaussian noise on weights or residual activations elicits hidden backdoor behaviors more often than temperature sampling on four of six models, with proxy-task hyperparameter selection via Thompson sampling improving results over uniform sweeps.
Anisotropy, quantified by dominant-dimension variance fraction, determines the best parameter-free similarity metric for text embeddings, with rank-based metrics gaining ~20% relative where cosine is weakest.
UCOB uses local return comparisons between skill and no-skill views to choose which view teaches the other, improving agent training on ALFWorld and WebShop.
CRAFT is a three-pillar credit assignment scheme that uses counterfactual token importance from GRPO sibling rollouts to provide signed per-token distillation signals in self-distilled agentic RL.
citing papers explorer
-
Expander Sparse Autoencoders: Parameter-Efficient Dictionaries for Mechanistic Interpretability
Expander SAEs apply left-d-regular expander masks to TopK SAEs, learning only dn decoder parameters instead of mn and tracing a storage-fidelity frontier that reaches 293x compression with 84% retained performance on Qwen2.5-3B.
-
TW-LegalBench: Measuring Taiwanese Legal Understanding
TW-LegalBench evaluates 13 LLMs on over 30,000 Taiwanese legal tasks from exams and judgments, showing top models pass lawyer thresholds but struggle with exact statute citations.
-
Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness
Fine-tuning updates frequently stale activation monitors for language model safety while quantization does not, with degradation predictable and repairable via label-free realignment.
-
Acceptance Cards:A Four-Diagnostic Standard for Safe Fine-Tuning Defense Claims
Acceptance Cards is a new four-diagnostic standard for safe fine-tuning defense claims that requires statistical reliability, fresh semantic generalization, mechanism alignment, and cross-task transfer; under this protocol SafeLoRA fails the full-card pass on Gemma-2-2B-it.
-
FormalRewardBench: A Benchmark for Formal Theorem Proving Reward Models
FormalRewardBench is the first benchmark for reward models in formal theorem proving, consisting of 250 Lean 4 preference pairs that show frontier LLMs scoring 59.8% while specialized provers score only 24.4%.
-
Unifying Scientific Communication: Fine-Grained Correspondence Across Scientific Media
Creates the first benchmark dataset integrating papers, slides, videos, and presentations for evaluating AI models on fine-grained multimodal correspondences in science.
-
ArgBench: Benchmarking LLMs on Computational Argumentation Tasks
ArgBench unifies 33 existing datasets into a standardized benchmark for testing LLMs across 46 argumentation tasks and analyzes the impact of prompting techniques and model factors on performance.
-
RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks
RLCracker is a reinforcement learning attack that erases LLM watermarks at 98.5% success rate with minimal data and generalizes across ten schemes and multiple model sizes.
-
Large Language Diffusion Models
LLaDA is a scalable diffusion-based language model that matches autoregressive LLMs like LLaMA3 8B on tasks and surpasses GPT-4o on reversal poem completion.
-
A First-Principles Theory of Slow Thinking and Active Perception
Active lifting of data distributions via latent-sequence sampling and max-rate uncertainty reduction formally derives slow-thinking LLMs and places them on representation and sampler hierarchies that can be climbed.
-
DecompRL: Solving Harder Problems by Learning Modular Code Generation
DecompRL is an RL method that learns modular code decomposition for LLMs, enabling exponential candidate generation via recombination to solve harder coding problems with lower GPU cost.
-
Conditional Co-Ablation: Recovering Self-Repair Backups in Transformer Circuits
Conditional Co-Ablation recovers self-repair backup heads in transformers by scoring conditional ablation growth, raising ROC-AUC from 0.33 to 0.91 on the IOI circuit and transferring to induction across models.
-
Multi-Head Recurrent Memory Agents
The paper proposes Multi-Head Recurrent Memory (MHM) with a select-then-update strategy to improve memory retention in long-context recurrent agents.
-
Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use
Static SFT and RL training for tool-use agents leads to performance drops under open-world distributional shifts across perception, interaction, reasoning and internalization; perturbation-augmented fine-tuning is proposed as mitigation.
-
Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
Multi-agent LLMs classify USPTO reactions and write verified SMIRKS rules, expanding a reaction taxonomy from 68 to 14,073 classes and matching proprietary classifiers on held-out and out-of-distribution data.
-
Beyond Activation Alignment:The Alignment-Diversity Tradeoff in Task-Aware LLM Quantization
TASA improves task-aware mixed-precision LLM quantization by searching calibration data mixtures via gradient-trace alignment and aggregating perplexity plus reasoning sensitivity signals, enabling 3.5-bit models to match or beat 4-bit baselines with over 20-point gains on GSM8K.
-
HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment
HARC couples harmfulness and refusal directions at prompt and response positions, yielding the best robustness-capability-usability trade-off among major safety methods.
-
SEFORA: Student Essays with Feedback Corpus and LLM Feedback Evaluation Framework
Releases SEFORA corpus of instructor feedback on college writing and UniMatch evaluation showing no LLM configuration exceeds 0.4 F1 in matching instructor priorities.
-
FlexiSLM: A Dynamic and Controllable Frame Rate Spoken Language Model
FlexiSLM is the first spoken language model supporting dynamic and controllable frame rates on speech input and output, outperforming fixed-rate 7B models at high quality and enabling faster inference at lower rates like 6.25 Hz.
-
CORTEX: High-Quality Cross-Domain Organization of Web-Scale Corpora through Ontological Corpus Graph
Cortex uses an Ontological Corpus Graph to structure web-scale corpora, creating a refined 24.14B-token corpus and a new benchmark validated on eight LLMs.
-
Fuzzing Large Language Models to Elicit Hidden Behaviours
Fuzzing via Gaussian noise on weights or residual activations elicits hidden backdoor behaviors more often than temperature sampling on four of six models, with proxy-task hyperparameter selection via Thompson sampling improving results over uniform sweeps.
-
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings
Anisotropy, quantified by dominant-dimension variance fraction, determines the best parameter-free similarity metric for text embeddings, with rank-based metrics gaining ~20% relative where cosine is weakest.
-
UCOB: Learning to Utilize and Evolve Agentic Skills via Credit-Aware On-Policy Bidirectional Self-Distillation
UCOB uses local return comparisons between skill and no-skill views to choose which view teaches the other, improving agent training on ALFWorld and WebShop.
-
CRAFT: Counterfactual Credit Assignment from Free Sibling Rollouts for Self-Distilled Agentic Reinforcement Learning
CRAFT is a three-pillar credit assignment scheme that uses counterfactual token importance from GRPO sibling rollouts to provide signed per-token distillation signals in self-distilled agentic RL.
-
Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution
Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.
-
Ko-WideSearch: A Korean Breadth-Search Benchmark for Exhaustive Set Enumeration by Web Agents
Ko-WideSearch is a new Korean breadth-search benchmark spanning 16 categories and three difficulty tiers that evaluates web agents on full set membership plus per-item attributes, showing consistent gaps between set recovery and row completion.
-
Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge
LaViD distills LLM conceptual knowledge to vision models via LLM-generated MCQ soft labels, outperforming vision-language distillation baselines on fine-grained benchmarks while improving robustness on spurious correlation datasets.
-
Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents
The supersession gap in LLM agents—failing to use current facts and discard superseded ones—is a distinct failure not fixed by scale or memory size, but improvable via RL training on a new environment.
-
BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents
BiPACE improves LLM agent policy optimization by using bisimulation proxies from hidden states for step clustering and action-conditioned baselines for advantage estimation, raising success rates on ALFWorld, WebShop, and TextCraft.
-
How Reliable Is Your Jailbreak Judge? Calibration and Adversarial Robustness of Automated ASR Scoring
Automated judges for LLM jailbreak ASR show opposite calibration failures and low robustness, with LLM judges flipped by benign framing and classifiers vulnerable to white-box attacks.
-
Beyond Trajectory Imitation: Strategy-Guided Policy Optimization for LLM Reasoning
SGPO extracts strategies from strong-model responses, builds autonomous and guided trajectories, and applies token-level forward-KL distillation with adaptive weighting to outperform SFT and RL baselines by 2.2 points on math benchmarks.
-
Mind the Heads: Topological Representation Alignment for Multimodal LLMs
HeRA aligns least-aligned attention heads in MLLMs using an MKNN-based contrastive objective to preserve cross-modal topological structure, yielding gains on vision-centric tasks and reduced hallucinations across 18 benchmarks.
-
When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents
Hidden-state convergence at step 4 predicts behavioral consistency in LLM agents on QA tasks (r=-0.35 to -0.83), enabling AUROC 0.97 detection of inconsistent trajectories but not improving accuracy on harder benchmarks.
-
Interleaved Speech Language Models Latently Work In Text
Interleaved SLMs implicitly transcribe spoken words to text tokens in middle layers (top candidate for 77% of data) before predicting in text space and returning to speech.
-
First-Token Broadcasters: Mechanistic Origins of Language Identity and Distributed Robustness in Transformers
Introduces LIHA ablation to locate first-token broadcaster heads and provides causal evidence that instruction tuning localizes language identity circuits to early layers in transformers.
-
NL2Scratch: An Executable Benchmark and Evaluation for Block-Based Programming
NL2Scratch supplies an executable benchmark of 311,648 NL-Scratch pairs and the SAC metric, showing LLMs with high lexical F1 often fail semantic alignment on actions, conditions, and numbers.
-
Mat-Pref: Verifiable-Reward Training Improves Compositional Reasoning in Inorganic Materials
Mat-Pref benchmark shows GRPO after SFT lets Qwen3-8B reach 65-72% on compositional materials reasoning tasks, exceeding zero-shot 235B models on held-out structure families and cross-property transfer.
-
MedHal-Loc: Are "Explainable-by-Architecture" Medical Hallucination Detectors Faithful Localizers? A Localization Benchmark
MedHal-Loc benchmark shows KG-triple hallucination detectors localize errors no better than chance on controlled medical statements due to entity extraction limits, while NLI and consistency methods succeed above chance, and real hallucinations are mostly diffuse conclusion changes.
-
FinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming
FinRED creates an expert-validated benchmark and rubric for financial LLM safety that maps regulatory standards to specific threats and reduces critical false negatives in evaluation from 28 to 12.
-
User as Engram: Internalizing Per-User Memory as Local Parametric Edits
User facts are internalized as surgical local edits to a hash-keyed Engram memory table with reasoning skill held in a shared adapter, claimed to match LoRA recall, improve indirect reasoning 5.6x on average, and compose across users with 33,000x smaller footprint than per-user adapters.
-
GraphPO: Graph-based Policy Optimization for Reasoning Models
GraphPO represents reasoning rollouts as a DAG to merge semantically equivalent paths, share suffixes, and assign separate efficiency and correctness advantages for lower variance and better performance than chain or tree baselines.
-
Lost in a Single Vector: Improving Long-Document Retrieval with Chunk Evidence Aggregation
DICE aggregates independently encoded document chunks into a single vector to reduce evidence dilution in long-document dense retrieval, reporting gains on LongEmbed especially beyond 4k tokens.
-
Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose
SkillWeaver formalizes compositional skill routing for LLM agents and introduces SAD, which raises step-level decomposition accuracy from 51% to 67.7% on a new 300-query benchmark over 2209 real MCP skills.
-
FllumaOne: A Code-Native Multimodal CAD Dataset with Executable Programs and Kernel-Validated Feature Histories
FllumaOne releases 100,000 kernel-validated CAD models as executable Python programs with aligned multimodal data including feature histories and geometry exports.
-
Looking Is Not Picking: An Attention-Segment Account of Tool-Selection Failures in LLM Agents
Attention analysis shows that LLM tool selection failures occur at the readout/decision stage, not because the model fails to attend to the correct tool definition.
-
Small LLMs for Biomedical Claim Verification: Cost-Effective Fine-Tuning, Structural Dataset Shortcuts, and Cross-Domain Generalization
Fine-tuned Mistral-7B via QLoRA achieves up to 12% higher F1 than GPT-4o on biomedical claim verification with 1008 examples, identifies a structural shortcut in SciFact, and shows robust cross-domain transfer from sound data.
-
Rubric-Guided Self-Distillation: Post-Training Without Rubric Verifiers
RGSD distills rubric-conditioned teacher distributions into base policies token-by-token, matching GRPO rubric satisfaction on Qwen models with one rollout and zero verifier calls.
-
Which Models Are Our Models Built On? Auditing Invisible Dependencies in Modern LLMs
ModSleuth reconstructs dependency graphs from public artifacts for four LLM releases, recovering 1,060 source-verified dependencies and exposing license issues, train-evaluation coupling, and documentation gaps.
-
Categorical Prior Lock-in: Why In-Context Learning Fails for Structured Data
ICL in LLMs shows a sharp ceiling on categorical distributions for high-cardinality tabular data, failing to reproduce rare classes despite examples, while numerical fidelity improves.
-
Harnessing Routing Foresight for Micro-step-level MoE load balancing in RL Post-training
ForeMoE uses routing foresight from the rollout stage to enable micro-step load balancing in MoE RL post-training via a hierarchical planner and transfer engine, claiming up to 1.45x speedup on 64 GPUs.