arxiv arXiv cs.CL · 2d ago

Latent Personal Memory: Dynamic Soft Prompts for LLM Personalization

Latent Personal Memory (LPM) represents user-specific memories as a compact, persistent matrix of N latent slots. These slots are mapped via a shared cross-attention network into dynamic, input-conditioned soft prompts that are prepended to a frozen LLM. LPM outperforms LoRA and Prompt Tuning by up to 8.8% and 54.4% on PersonaMem v1, reduces KV-cache usage by over 64x, matches LoRA accuracy on LoCoMo with 120x fewer parameters, and scales efficiently with context length, outperforming full-context at 128K tokens.

arxiv arXiv cs.LG · 6d ago

Probe-and-Refine Tuning Improves Coding Agent Performance

A new method called probe-and-refine tuning uses synthetic bug-fix probes to iteratively improve repository guidance files with single-shot LLM calls, without agent loops or tool use. On SWE-bench Verified, it achieves a 33.0% mean resolve rate—14.5 percentage points higher than the initial static knowledge base—showing improved coverage rather than patch precision. The method enables agents to use larger step budgets effectively, and performance remains stable across models when diagnostic output is sufficient.

arxiv arXiv cs.AI · 6d ago

SoftSkill: Behavioral Compression for Contextual Adaptation

SoftSkill proposes a method to compress natural-language skills into compact latent priors, improving task performance on SearchQA, LiveMath, and DocVQA. It outperforms SkillOpt by 5.2 to 12.5 points on key benchmarks while replacing hundreds to thousands of Markdown tokens with a few virtual tokens.

arxiv arXiv cs.CL · 6d ago

AgentFinVQA: Auditable, On-Premise Financial Chart QA

AgentFinVQA introduces a multi-agent pipeline for financial chart question answering that ensures auditability and on-premise deployability without significant accuracy loss. It outperforms baseline models by +7.68 pp using a proprietary backbone and +4.84 pp with open-weights Qwen3.6-27B-FP8, while providing a confidence signal via verifier output that improves human review routing.

arxiv arXiv cs.CL · 6d ago

Selective Verification for Budget-Aware Reasoning

Sevra, a serving-layer controller, selectively verifies answers to improve accuracy and reduce token usage. On \mathfive, it achieves 76.3% accuracy with 26.8% fewer post-generation tokens and halved harmful flips, while on \gsm it verifies only 3.0% of examples, boosting accuracy to 94.5% and cutting verification tokens by 91.2%. The study shows that initial solve length and explicit control needs determine optimal verification strategy.

arxiv arXiv cs.CL · 6d ago

STAGE: Source-Grounded Data Generation for Text-to-JSON

STAGE is a pipeline that generates text-to-JSON training data by using LLMs to synthesize reports and JSON schemas, validated against underlying spreadsheets. Evaluations on STAGE-Eval show it improves Qwen3-4B exact match from 31.37% to 74.27% and value accuracy from 45.46% to 90.69%.

arxiv arXiv cs.CL · 6d ago

HydraHead: Head-Level Hybrid Attention for Long-Context Performance

HydraHead introduces a head-level hybridization of Full and Linear Attention, leveraging interpretability to select retrieval-critical heads and fuse outputs via a scale-normalized module. Trained on 15B tokens, it achieves over 69% improvement over baseline at 512K context length, outperforming layer-wise hybrids and approaching Qwen3.5's performance on long-context tasks.

arxiv arXiv cs.CL · 6d ago

Causal Activation Directions for Mitigating Emergent Misalignment in Language Models

Fine-tuning language models on insecure code causes emergent misalignment. A shared activation direction across four model families achieves 99.6% separation of aligned and misaligned activations, and subtracting it reduces code spillover by 21-51 points. Cross-architecture transfer shows behavioral suppression but lacks specificity, with within-model directions being causally actionable and cross-model directions only causally real.

arxiv arXiv cs.LG · 7d ago

MAST Enables Selective Unlearning in RLVR-Induced Reasoning

MAST, a mechanism-guided unlearning method, achieves targeted forgetting of RLVR-induced reasoning with minimal collateral damage. On Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, it significantly reduces MATH performance (45/150 to 37/15-0) while preserving GSM8K accuracy by +0.8 points and maintaining MATH retention at -0.5 points. Results hold across different seeds, objectives, and models, showing superior stability over full-parameter unlearning.

arxiv arXiv cs.AI · 7d ago

MAST Enables Selective Unlearning in RLVR-Induced Reasoning

MAST, a mechanism-guided unlearning method, achieves targeted forgetting of RLVR-induced reasoning with minimal collateral damage. On Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, it significantly reduces MATH performance (45/150 to 37/15-0) while preserving GSM8K accuracy by +0.8 points and maintaining MATH retention at -0.5 points. Results hold across seeds, objectives, and models, showing superior stability over full-parameter unlearning.

arxiv arXiv cs.CL · 7d ago

OmniAgent: Native Active Perception for Omni-Modal Understanding

OmniAgent introduces a POMDP-based iterative Observation-Thought-Action cycle for video understanding, enabling on-demand action execution to selectively distill audio-visual cues into persistent textual memory. It achieves state-of-the-art performance on ten benchmarks, with a 7B agent outperforming a 10× larger Qwen2.5-VL-72B model on LVBench (50.5% vs. 47.3%).

arxiv arXiv cs.LG · 7d ago

Spotlight: Using Spot GPUs to Accelerate DiT RL Post-Training

Spotlight enables DiT RL post-training by leveraging idle spot GPUs, reducing costs by 1.4-6.4x while achieving superior image quality. It uses stale model weights in exploration and reconfigures sequence parallelism on-the-fly, allowing efficient GPU utilization without breaking training pipelines.

arxiv arXiv cs.CL · 7d ago

PragReST: Self-Reinforcing Counterfactual Reasoning for Pragmatic Language Understanding

PragReST is a self-supervised framework that enhances large language models' pragmatic reasoning by generating counterfactual reasoning traces and training via supervised fine-tuning and reinforcement learning. It outperforms baseline models on four pragmatic benchmarks, improving Qwen3-8B and Qwen3-14B by 5.37% and 5-5.50% accuracy respectively, and maintains strong performance on general-knowledge and mathematical reasoning tasks.

arxiv arXiv cs.CL · 7d ago

Data Recipe Boosts Long-Context Reasoning in LLMs

A data-centric approach improves long-context reasoning in large language models, using eight curated datasets with 14K examples across retrieval, multi-evidence synthesis, and reasoning tasks. When paired with minimal outcome-based GRPO training, it achieves average gains of +7.2 to +6.4 points on seven benchmarks, outperforming prior RL training sets, and enhances agentic performance by +4.8 and +7.0 points on GAIA and BrowseComp respectively.

arxiv arXiv cs.AI · 7d ago

Data Recipe Boosts Long-Context Reasoning in LLMs

A data-centric approach improves long-context reasoning in large language models, using eight curated datasets with 14K examples across retrieval, multi-evidence synthesis, and reasoning tasks. When paired with minimal outcome-based GRPO training, it achieves average gains of +7.2 to +6.4 points on seven benchmarks, outperforming prior RL training sets, and enhances agentic performance by +4.8 and +7.0 points on GAIA and BrowseComp respectively.

arxiv arXiv cs.AI · 7d ago

Spotlight: Using Spot GPUs to Accelerate DiT RL Post-Training

Spotlight enables DiT RL post-training by leveraging idle spot GPUs, reducing costs by 1.4-6.4× while achieving superior image quality. It uses stale model weights in exploration and reconfigures sequence parallelism in real time, allowing efficient GPU utilization without breaking training pipelines.

arxiv arXiv cs.CL · 8d ago

Geographic Bias in Large Language Models from User Metadata

A study reveals that even neutral prompts trigger region-specific responses in large language models due to user metadata. Location leakage increases by up to 793 times in some models, and using 'Unknown' instead of location metadata still causes significant bias, indicating the user profile frame itself acts as a conditioning signal.

arxiv arXiv cs.CL · 8d ago

RubricsTree: Scalable Evaluation Framework for Personal Health Agents

RubricsTree introduces a hierarchical taxonomy of over 100 clinically-verifiable Boolean rubrics, evolved from 4,000 real user queries via human-in-the-loop curation. It enables scalable, expert-aligned evaluation of personal health agents by dynamically routing queries to relevant rubrics and outperforms baseline methods in alignment, context sensitivity, and model performance gains of up to 66% on HealthBench.

arxiv arXiv cs.CL · 8d ago

ZPPO: Teacher in Prompts, Not Gradients

Zone of Proximal Policy Optimization (ZPPO) integrates teacher knowledge directly into prompts rather than policy gradients. It uses Binary and Negative Candidate-included Questions to surface student failure modes and amplifies learning through a prompt replay buffer, achieving superior performance on hard questions across student scales, especially at smaller model sizes.

arxiv arXiv cs.LG · 8d ago

EnvRL: Leveraging Environment Dynamics in Agentic RL

EnvRL introduces a framework that enhances agentic reinforcement learning by incorporating environment dynamics through state prediction and inverse dynamics objectives. When trained with GRPO, EnvRL improves success rates of Qwen-2.5-1.5B-Instruct from 72.8% to 77.4% on ALFWorld and from 56.8% to 67.0% on WebShop.