Maven is an RL method using answer-conditioned evidence-state values to assign rewards to add, link, and drop actions on evidence memory, outperforming outcome-only baselines on LongBench v2, LongReason, and RULER.
Title resolution pending
105 Pith papers cite this work. Polarity classification is still indexing.
105
Pith papers citing it
citation-role summary
background 3
citation-polarity summary
roles
background 3representative citing papers
Pre-training 6B LLMs on temporally ordered Common Crawl snapshots yields models with improved factual freshness and temporal precision over shuffled baselines while matching on general language understanding.
Proves that RoPE attention loses locality bias and token distinction in long contexts, approaching random behavior independent of content.
DiHAL uses geometry proxies to pick where to replace the lower layers of a pretrained transformer with a diffusion bridge for hidden-state reconstruction, improving over token-level diffusion baselines on 8B models.
LLMs can provide cost-effective annotation of credibility in Danish asylum texts but produce inconsistent errors that vary by model and prompt, requiring checks beyond single-model accuracy.
AAAC learns two 64-byte codebooks per layer for 4-bit LLM weights and lets each group pick the one minimizing activation-weighted reconstruction error, storing the choice at zero extra cost.
Language model circuits show high within-task consistency and necessity but substantial overlap across tasks, making them less specific than assumed.
Energy-navigated trajectory shaping during training produces 8-step discrete flow matching students that achieve 32% lower perplexity than 1024-step teachers on 170M language models with unchanged inference cost.
Introduces the GeoDial dataset of 1.3K multimodal geometry tutoring dialogs grounded in diagram highlights, proposes an annotation protocol, and shows that fine-tuned VLMs improve dialog but struggle with accurate highlights.
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.
CircuitFormer is a 511M-parameter encoder-decoder model that generates analog circuit topologies from text prompts at 100% syntactic correctness and 83% functional success using a new subcircuit-mining tokenizer that keeps vocabulary size fixed at 512.
LLMs process negation using both attention-based suppression and constructive representation mechanisms (construction dominant), with late-layer attention shortcuts explaining poor accuracy on negation tasks.
Surprisal minimization over goal-directed alternatives generated by language models provides the strongest account of production choices in open-ended dialogue compared to uniform information density or length-based costs.
BWLA is the first post-training quantization method for LLMs that achieves 1-bit weights paired with low-bit activations such as 6 bits, using OKT to reshape weights and suppress activation tails plus PSP for low-rank refinement.
LLMs display inconsistent factual recall across different surface forms of the same entity, with greater robustness to minor spelling changes than to aliases or abbreviations.
Adaptive Instruction Composition uses a neural contextual bandit with RL to adaptively combine crowdsourced texts, generating more effective and diverse LLM jailbreaks than random or prior adaptive methods on Harmbench.
Uncertainty and correctness in LLMs are encoded by distinct feature populations, with suppression of confounded features improving accuracy and reducing entropy.
MUCOCO applies semantic-preserving mutation analysis to automatically expose inconsistent behaviors in code LLMs, detecting inconsistencies in about 15% of cases across 7 models and 4 tasks while outperforming the TURBULENCE baseline.
STAR-Teaming uses a Strategy-Response Multiplex Network inside a multi-agent framework to organize attack strategies into semantic communities, delivering higher attack success rates on LLMs at lower computational cost than prior methods.
R-CAI inverts constitutional AI to automatically generate diverse toxic data for LLM red teaming, with probability clamping improving output coherence by 15% while preserving adversarial strength.
VLMs as judges exhibit informativeness bias by favoring detailed but image-inconsistent answers; BIRCH mitigates it by first correcting answers against the image, reducing bias up to 17% and improving performance up to 9.8%.
CODI compresses explicit CoT into continuous space via self-distillation and is the first implicit method to match explicit CoT performance on GSM8k at GPT-2 scale with 3.1x compression and 28.2% higher accuracy than prior implicit approaches.
SWE-RL uses RL on software evolution data to train LLMs achieving 41% on SWE-bench Verified with generalization to other reasoning tasks.
Omni-MATH supplies 4428 human-verified Olympiad math problems that expose top LLMs achieving only 52.55% to 60.54% accuracy on the most difficult items.
citing papers explorer
-
Evidence-State Rewards for Long-Context Reasoning
Maven is an RL method using answer-conditioned evidence-state values to assign rewards to add, link, and drop actions on evidence memory, outperforming outcome-only baselines on LongBench v2, LongReason, and RULER.
-
Understanding Data Temporality Impact on Large Language Models Pre-training
Pre-training 6B LLMs on temporally ordered Common Crawl snapshots yields models with improved factual freshness and temporal precision over shuffled baselines while matching on general language understanding.
-
RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably
Proves that RoPE attention loses locality bias and token distinction in long contexts, approaching random behavior independent of content.
-
Where Should Diffusion Enter a Language Model? Geometry-Guided Hidden-State Replacement
DiHAL uses geometry proxies to pick where to replace the lower layers of a pretrained transformer with a diffusion bridge for hidden-state reconstruction, improving over token-level diffusion baselines on 8B models.
-
LLMs as annotators of credibility assessment in Danish asylum decisions: evaluating classification performance and errors beyond aggregated metrics
LLMs can provide cost-effective annotation of credibility in Danish asylum texts but produce inconsistent errors that vary by model and prompt, requiring checks beyond single-model accuracy.
-
AAAC: Activation-Aware Adaptive Codebooks for 4-bit LLM Weight Quantization
AAAC learns two 64-byte codebooks per layer for 4-bit LLM weights and lets each group pick the one minimizing activation-weighted reconstruction error, storing the choice at zero extra cost.
-
How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits
Language model circuits show high within-task consistency and necessity but substantial overlap across tasks, making them less specific than assumed.
-
Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation
Energy-navigated trajectory shaping during training produces 8-step discrete flow matching students that achieve 32% lower perplexity than 1024-step teachers on 170M language models with unchanged inference cost.
-
GeoDial: A Multimodal Conversational Tutoring Dataset for Geometry Problem-Solving with Visual Tutor Turns
Introduces the GeoDial dataset of 1.3K multimodal geometry tutoring dialogs grounded in diagram highlights, proposes an annotation protocol, and shows that fine-tuned VLMs improve dialog but struggle with accurate highlights.
-
Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost
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.
-
CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt
CircuitFormer is a 511M-parameter encoder-decoder model that generates analog circuit topologies from text prompts at 100% syntactic correctness and 83% functional success using a new subcircuit-mining tokenizer that keeps vocabulary size fixed at 512.
-
How Language Models Process Negation
LLMs process negation using both attention-based suppression and constructive representation mechanisms (construction dominant), with late-layer attention shortcuts explaining poor accuracy on negation tasks.
-
Surprisal Minimisation over Goal-directed Alternatives Predicts Production Choice in Dialogue
Surprisal minimization over goal-directed alternatives generated by language models provides the strongest account of production choices in open-ended dialogue compared to uniform information density or length-based costs.
-
BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs
BWLA is the first post-training quantization method for LLMs that achieves 1-bit weights paired with low-bit activations such as 6 bits, using OKT to reshape weights and suppress activation tails plus PSP for low-rank refinement.
-
Revisiting Non-Verbatim Memorization in Large Language Models: The Role of Entity Surface Forms
LLMs display inconsistent factual recall across different surface forms of the same entity, with greater robustness to minor spelling changes than to aliases or abbreviations.
-
Adaptive Instruction Composition for Automated LLM Red-Teaming
Adaptive Instruction Composition uses a neural contextual bandit with RL to adaptively combine crowdsourced texts, generating more effective and diverse LLM jailbreaks than random or prior adaptive methods on Harmbench.
-
Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders
Uncertainty and correctness in LLMs are encoded by distinct feature populations, with suppression of confounded features improving accuracy and reducing entropy.
-
MUCOCO: Automated Consistency Testing of Code LLMs
MUCOCO applies semantic-preserving mutation analysis to automatically expose inconsistent behaviors in code LLMs, detecting inconsistencies in about 15% of cases across 7 models and 4 tasks while outperforming the TURBULENCE baseline.
-
STAR-Teaming: A Strategy-Response Multiplex Network Approach to Automated LLM Red Teaming
STAR-Teaming uses a Strategy-Response Multiplex Network inside a multi-agent framework to organize attack strategies into semantic communities, delivering higher attack success rates on LLMs at lower computational cost than prior methods.
-
Reverse Constitutional AI: A Framework for Controllable Toxic Data Generation via Probability-Clamped RLAIF
R-CAI inverts constitutional AI to automatically generate diverse toxic data for LLM red teaming, with probability clamping improving output coherence by 15% while preserving adversarial strength.
-
When Vision-Language Models Judge Without Seeing: Exposing Informativeness Bias
VLMs as judges exhibit informativeness bias by favoring detailed but image-inconsistent answers; BIRCH mitigates it by first correcting answers against the image, reducing bias up to 17% and improving performance up to 9.8%.
-
CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation
CODI compresses explicit CoT into continuous space via self-distillation and is the first implicit method to match explicit CoT performance on GSM8k at GPT-2 scale with 3.1x compression and 28.2% higher accuracy than prior implicit approaches.
-
SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution
SWE-RL uses RL on software evolution data to train LLMs achieving 41% on SWE-bench Verified with generalization to other reasoning tasks.
-
Omni-MATH: A Universal Olympiad Level Mathematic Benchmark For Large Language Models
Omni-MATH supplies 4428 human-verified Olympiad math problems that expose top LLMs achieving only 52.55% to 60.54% accuracy on the most difficult items.
-
It Takes a MAESTRO To Prune Bad Experts
Pruning MoE LLMs according to the stationary distribution of a Markov chain over (layer, expert) routing transitions retains more task performance than local importance heuristics, with up to ~3.5% relative gains over the strongest baseline at 50% compression.
-
DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation
Dialectal robustness and generation are dissociated in LLMs: benchmarks are driven by pretraining and SFT while alignment reshapes generation invisibly to benchmarks, and the method maximizing dialectal reward is least preferred by human evaluators.
-
MemDefrag: Latent Memory Defragmentation for Large Language Models
Ranking latent-memory fragments by middle-layer attention density and keeping only the top few more than doubles long-term knowledge-retention accuracy (43.0% vs. 17.4%/17.6%) versus MemoryLLM and M+.
-
How Much is Left? LLMs Linearly Encode Their Remaining Output Length
Linear probes on frozen LLM hidden states recover an approximate remaining-output-length signal that is decodable at prompt-end, transfers across datasets, and shifts upward at retraction tokens.
-
Understanding Why Language Models Hallucinate: Testing Reasoning Against Priors
Hallucinations arise from biased latent inference paths rather than missing knowledge, demonstrated via a new diagnostic testbed TrapQA that isolates task-retrieval and key-selection biases.
-
SeKV: Resolution-Adaptive KV Cache with Hierarchical Semantic Memory for Long-Context LLM Inference
SeKV introduces resolution-adaptive semantic KV caching with GPU-CPU hierarchy and selective zoom-in reconstruction, achieving 5.9% average improvement over semantic baselines and 53.3% GPU memory reduction at 128K context.
-
When Calibration Rankings Reverse: Accuracy-Controlled Evaluation for Fair Comparison of LLMs
Global calibration metrics like ECE are confounded by accuracy; the proposed ACE framework with three accuracy-controlled views shows many prior calibration advantages weaken or reverse.
-
The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning
Benign multilingual fine-tuning causes language-specific safety drifts with adversarial compliance rates rising up to four-fold, decoupled from capability gains.
-
Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models
Quantized reasoning models produce longer chains of thought, inflating token usage and negating per-token speedups from low-bit quantization across multiple benchmarks.
-
Tail-Shape Estimation in LLM Evaluation Is Fragile: A Protocol for Diagnosing False Positives
A pre-registered protocol rejects tail-shape claims in standard LLM toxicity evaluation as false positives under two different scorer families.
-
PGT: Procedurally Generated Tasks for improving visual grounding in MLLMs
PGT generates synthetic tasks via geometric overlays on images to supply dense visual supervision, improving spatial and relational understanding in MLLMs by up to 20% on targeted benchmarks.
-
Ex-GraphRAG: Interpretable Evidence Routing for Graph-Augmented LLMs
Ex-GraphRAG replaces GNN encoders with M-GNAN for exact node-level decomposition in graph-augmented LLMs, matching black-box performance on STaRK-Prime while exposing semantic-structural mismatches that degrade multi-hop QA when low-attribution intermediaries are removed.
-
How Far Will They Go? Red-Teaming Online Influence with Large Language Models
An empirical red-teaming study measures political Overton Windows across more than 30 open-source LLMs from 10 families and finds left-leaning bias, inverse size correlation, regional variation, and variable jailbreak effectiveness.
-
How Much Online RL is Enough? Informative Rollouts for Offline Preference Optimization in RLVR
Short GRPO warm-up followed by offline DPO on informative rollouts matches or beats full GRPO on math reasoning benchmarks at substantially lower compute cost.
-
BLINKG: A Benchmark for LLM-Integrated Knowledge Graph Generation
BLINKG is a benchmark for evaluating LLMs on mapping input data schemas to ontology concepts for knowledge graph construction, with experiments showing promising but limited performance in complex real-world scenarios.
-
Forecasting Downstream Performance of LLMs With Proxy Metrics
Proxy metrics from next-token distributions over expert solutions outperform loss and compute baselines for ranking LLMs, selecting pretraining data, and extrapolating performance across compute scales.
-
DMN: A Compositional Framework for Jailbreaking Multimodal LLMs with Multi-Image Inputs
DMN achieves over 90% attack success rate on GPT-4o, Gemini-2.5-pro and Claude Sonnet 4 by distributing instructions, supplying multimodal evidence, and adding number chain tasks across multiple images.
-
Self-Supervised On-Policy Distillation for Reasoning Language Models
SSOPD converts intra-group correct-wrong contrast into process supervision by distilling a teacher distribution from the shortest correct completion into prefixes of the longest wrong completion, improving GRPO on AIME and HMMT benchmarks.
-
MixSD: Mixed Contextual Self-Distillation for Knowledge Injection
MixSD uses dynamic mixing of the model's expert and naive conditionals to create distribution-aligned supervision that improves the memorization-retention tradeoff over standard SFT.
-
GRLO: Towards Generalizable Reinforcement Learning in Open-Ended Environments from Zero
GRLO shows RLHF from scratch on 5K open-ended prompts raises average performance from 24.1 to 63.1 across domains on Qwen3-4B-Base using 46x less data and 68x less compute than in-domain RLVR while remaining competitive with heavily post-trained models.
-
From Text to Voice: A Reproducible and Verifiable Framework for Evaluating Tool Calling LLM Agents
A dataset-agnostic framework converts text tool-calling benchmarks to paired audio evaluations via TTS, speaker variation and noise, then evaluates seven omni-modal models showing model- and task-dependent performance with small text-to-voice gaps.
-
Polar probe linearly decodes semantic structures from LLMs
LLMs represent semantic relations geometrically via embedding distance and direction; a linear Polar Probe decodes these structures from middle-layer activations and generalizes to new entities.
-
Dual Hierarchical Dialogue Policy Learning for Legal Inquisitive Conversational Agents
A dual hierarchical RL framework with two agents coordinates high-level dialogue strategy and low-level question generation to emulate judicial questioning and extract key information from Supreme Court arguments, outperforming baselines.
-
Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time
OP-Mix is an on-policy data mixing method that uses low-rank adapter interpolation to find near-optimal data mixtures throughout language model training with reduced compute.
-
Low-Rank Adapters Initialization via Gradient Surgery for Continual Learning
SLICE applies gradient surgery via projection and truncated SVD to initialize LoRA adapters, yielding better stability-plasticity trade-offs on continual learning benchmarks including adversarial task sequences.
-
ODRPO: Ordinal Decompositions of Discrete Rewards for Robust Policy Optimization
ODRPO decomposes discrete rewards into ordinal binary indicators to create robust, variance-aware advantage estimators for noisy RLAIF in LLM alignment.