SignRoundV2 is a post-training quantization framework for LLMs maintaining performance under extreme compression (2-4 bits). It combines adaptive mixed-precision strategy guided by gradients and lightweight stabilization techniques. Results show ~1% performance gap at 4.5 bits average in mixed MXFP, with substantial improvements in challenging 2-bit weight-only quantization.
arXiv paper introducing a trace-optional evaluation protocol decomposing token efficiency of reasoning LLMs. Analyzes 14 open-weight models on CogniLoad, GSM8K, ProofWriter, ZebraLogic by separating completion rate, conditional correctness, and generated length. Identifies three failure modes: logic-limited, context-limited, or verbosity-limited.
EnvFactory automates creation of executable environments and synthesis of multi-turn trajectories for Agentic RL training. Using 85 verified environments across 7 domains, the framework generates 2,575 SFT/RL trajectories and improves Qwen3-series models by +15% on BFCLv3, +8.6% on MCP-Atlas, and +6% on conversational benchmarks.
History-Echoes framework investigates how conversational history biases LLM outputs. Using Markov chain modeling and geometric analysis of hidden representations across three model families and six datasets, researchers show behavioral persistence manifests as a geometric trap constraining model trajectories. Code released.
SDRL (Self-Debate Reinforcement Learning) trains LLMs to solve problems standalone AND benefit from multi-agent debate. The model samples multiple solutions, constructs debate context with diverse reasoning paths, then jointly optimizes initial and debate-conditioned responses. Results: consistent MAD performance gains across benchmarks and agent configurations.
OptimusVLA, a hierarchical Vision-Language-Action model, improves robotic manipulation via two memories: Global Prior Memory (replaces Gaussian noise with trajectory priors) and Local Consistency Memory (enforces temporal coherence). Results: 98.6% on LIBERO, +13.5% vs pi_0 on CALVIN, 2.9x inference speedup.
Med-V1 is a family of 3-billion-parameter language models trained on synthetic data for biomedical evidence attribution. It outperforms base models by +27% to +71% on five benchmarks and matches frontier LLMs like GPT-5, while detecting hallucinations and misattributions in clinical guidelines.
SkillMOO optimizes skill bundles for coding agents via multi-objective search (NSGA-II) on pass rate and inference cost. Across 16 SkillsBench SE tasks, the framework achieves top pass rate rank on 11/12 non-zero-pass tasks, reducing costs up to 31.7% and gaining up to 21 percentage points in pass rate.
SynCABEL uses LLMs to generate contextualized synthetic training data to address the scarcity of expert annotations in biomedical entity linking. The framework achieves SOTA results on MedMentions (English), QUAERO (French), and SPACCC (Spanish), reaching full human supervision performance with 60% less annotated data.
Theoretical study proving multi-layer cross-attention is optimal for multi-modal in-context learning. Authors show single-layer linear self-attention fails to recover Bayes-optimal predictor, but linearized cross-attention mechanism achieves Bayes optimality with gradient flow.
Preregistered study comparing Vector RAG and LLM-compiled markdown wiki on 13 questions over 24 papers. Wiki excels at cross-paper synthesis and claim-level citation accuracy, but uses more query tokens. A decomposition-based RAG variant recovers most wiki advantages at lower LLM-token cost.
OmniCode is a benchmark for evaluating AI agents on software engineering tasks. It contains 1794 tasks across Python, Java, and C++ covering bug fixing, test generation, code review fixing, and style fixing. Evaluations show SWE-Agent achieves only 25% on C++ test generation with DeepSeek-V3.1.
UbuntuGuard is the first policy-based safety benchmark for African languages. Built from 155 domain experts, it evaluates 15 models (7 general-purpose LLMs, 8 guardian models) across three variants. Findings show English-centric benchmarks overestimate real-world multilingual safety and cross-lingual transfer remains insufficient.
BELIEF combines structured evidence modeling and uncertainty-aware fusion for biomedical question answering. The framework converts retrieved documents into evidence objects (clinical attributes, source quality, relevance, support strength) and fuses two reasoning paths: symbolic (Dempster-Shafer theory) and neural (LLM). SOTA results on PubMedQA, MedQA, MedMCQA across 5 LLM backbones.
ACToR, an adversarial multi-agent loop, improves C-to-Rust translation using a translator agent and discriminator agent that iteratively compete. On 63 real-world C utilities (avg 473 LOC), the system achieves 90% test pass rate with zero human intervention, improving correctness by 36.7% over non-adversarial baselines.
SD-Search introduces on-policy hindsight self-distillation for search-augmented reasoning agents. A single model acts as both student and teacher: the teacher, conditioned on past rollout outcomes, guides the student via token-level Jensen-Shannon divergence at query positions. No external teacher model or additional annotations needed.
TaskGround is a task-inference framework for household agents operating on complete scenes. It structures reasoning in three steps: grounding (extracting relevant context), inference (executable structure), execution (action sequences). Evaluated on FullHome (400 tasks), it improves success rates and makes Qwen3.5-9B competitive with GPT-4 while reducing token costs by 18x.
LoopQ is a post-training quantization (PTQ) framework for recursive language models (LoopLMs) that reuse Transformer blocks. It addresses three challenges: distribution shift across roles, state reuse between loops, and recursive error accumulation. Under W4A4 quantization, LoopQ improves downstream accuracy by 68.8% and reduces perplexity by 87.7% versus static PTQ baseline.
R&B-EnCoRe enables Vision-Language-Action models to self-generate and refine embodied reasoning without human annotation or external rewards. Tested on manipulation (Franka Panda, WidowX), navigation (bipedal, wheeled, bicycle, quadruped), and autonomous driving: +28% manipulation success, +101% navigation scores, −21% collision rate vs baselines.
SocialMemBench evaluates AI memory systems in multi-party social group settings. The benchmark includes 430 personas, 7,355 conversation turns, and 1,031 QA pairs across 5 social archetypes. Gemini 2.5 Flash reaches 0.721 on small networks; open-source frameworks (Mem0, LangMem, Graphiti, Cognee) plateau at 0.12-0.18, revealing a significant capability gap.
Colluding LLM agents manipulate victim beliefs by coordinating truthful evidence fragments through public channels without covert communication. The Generative Montage framework (Writer-Editor-Director) constructs deceptive narratives via adversarial debate. Attack success rates reach 74.4% on proprietary models and 70.6% on open-weights across 14 LLM families. Advanced reasoning models show higher susceptibility.
SignRoundV2 is a post-training quantization framework for LLMs maintaining performance under extreme compression (2-4 bits). It combines adaptive mixed-precision strategy guided by gradient information and lightweight stabilization techniques. Results show ~1% performance gap at 4.5 bits average in mixed MXFP settings, with substantial improvements in 2-bit weight-only quantization.
New approach to generate formal theorem proving challenges by leveraging theoretical computer science (TCS). Framework automatically synthesizes problem-proof pairs in Lean4 and Markdown across two domains: Busy Beaver and Mixed Boolean Arithmetic. DeepSeekProver-V2-671B achieves 57.5% on Busy Beaver but only 12% on Mixed Boolean Arithmetic, revealing major gaps in long-form proof generation.
LGBO (LLM-Guided Bayesian Optimization) embeds LLM semantic reasoning into every Bayesian Optimization iteration via a region-lifted preference mechanism. Tested on physics, chemistry, biology, and materials science benchmarks, LGBO reaches 90% of best observed value in 6 iterations for Fe-Cr battery electrolyte optimization, versus 10+ for standard BO.
DISA is an offline RL method for LLMs that decouples partition-function estimation (via importance sampling) from policy optimization. On 9 benchmarks (math and code), it matches or exceeds FlowRL, outperforms GRPO/GSPO, and retains substantially more strategy-level diversity than reward-maximization baselines.
Fourier Compressor compresses visual tokens in Vision-Language Models using Fourier transforms. The parameter-free method reduces FLOPs by 83.8% and boosts inference speed by 31.2% while retaining 96% of original accuracy. Tested on LLaVA and Qwen-VL, it generalizes to video understanding tasks.
Perovskite-R1 is a specialized LLM based on QwQ-32B, fine-tuned on 1,232 scientific publications and 33,269 candidate materials to discover precursor additives optimizing perovskite solar cells. The model generates solutions for defect passivation and improves stability/performance, experimentally validated.
CoDiLA combines diffusion with local autoregressive decoding for parallel code generation. A compact auxiliary AR model (0.6B parameters) ensures fine-grained syntactic coherence while the diffusion model handles bidirectional block-level generation. Eliminates coherence artifacts on code generation benchmarks.
Formal verification framework for LLM pipelines using Lean 4 certificates. Three certificate families (conflict-aware bilattice, embedding sensitivity, Hoare-style agent action) plus two operators (Maximal Certifiable Residue, Compositional Stability) for high-stakes deployments (regulated finance, clinical support, agentic systems). Compiled artifact covers 22 certificate types, 17/46 declarations axiom-free.
Complete first-order analysis of gradient dynamics in transformer attention heads under cross-entropy training. Authors establish an advantage-based routing law and responsibility-weighted value updates, showing that optimization creates Bayesian manifolds implementing in-context probabilistic reasoning.
RAP is an elastic pruning framework for LLM inference using reinforcement learning to dynamically adapt compression strategies based on runtime memory variations and heterogeneous KV-cache demands. The RL agent optimizes the parameter-to-KV-cache ratio in real-time, retaining only components that maximize utility within the current memory budget.
DriveMoE proposes a Mixture-of-Experts architecture for end-to-end autonomous driving. The model combines Vision MoE (dynamic camera selection based on driving context) and Action MoE (specialized expert activation for different behaviors). Built on Drive-π₀ baseline, DriveMoE achieves SOTA on Bench2Drive by avoiding mode averaging.
InvDesFlow-AL combines diffusion and active learning for inverse design of functional materials. The model achieves RMSE 0.0423 Å in crystal structure prediction (+32.96% vs existing methods) and systematically generates low-formation-energy materials. Validation: discovery of Li₂AuH₆ as BCS superconductor at 140 K.
KISS introduces a Knowledge Infrastructure (KI) enabling AI agents to execute complex Earth science simulations. On 3,000 trials, KI-equipped agents produced physically plausible simulations in 84% of cases vs. <40% without KI. An automated Knowledge Dissection Toolkit (KDT) generated 119 KIs across 14 Earth-science domains, showing operational expertise is structured and extractable rather than ad hoc.
ReTAMamba proposes a Mamba-based architecture for predicting irregular clinical time series. The model estimates observation reliability from missingness and elapsed time, integrates short/long-term information via Chronological Weaving, and uses a budgeted token router. On MIMIC-IV, eICU, and PhysioNet 2012, AUPRC gains of 7.51%, 7.80%, and 10.15% respectively.
Amazon Music deploys a robust neural sparse retrieval system for large-scale music search. The system handles misspellings, transpositions, and phonetic variations with 91.4% recall@10 on 6M documents, outperforming trigrams (57.7%). Inference-free architecture with granular subword tokenization (max 3 chars) and zero online latency.
LiRA, a lightweight fine-tuning framework, improves LLM adaptation to low-resource languages through cross-lingual semantic alignment. Combines Arca (anchor-based alignment to English) and LaSR (language-aware head). Theoretical stability guarantees. Multilingual dataset (7 Asian languages) and code released.
RAG-Gym, a framework supervising the search process rather than final answers, improves autonomous search agents. Re²Search++, a process-supervised agent, achieves substantial gains on multi-hop information-seeking benchmarks, especially out-of-domain, through higher-quality search queries and better generalization.
SD-Search introduces on-policy hindsight self-distillation for search-augmented reasoning agents. A single model acts as both student (inference-time context only) and teacher (conditioned on search outcomes from rollout groups). Step-level supervision via Jensen-Shannon divergence at query positions, integrated into GRPO training without external models or annotations.
QuNetQFL is a quantum federated learning protocol implemented on quantum networks, masking local model updates with distributed quantum secret keys for information-theoretic security. Experimentally validated on a four-client quantum network, it reduces communication costs by 75% and scales to 200 clients with rapid convergence.