Loyola's TLA-Prover Beats Untuned AI at Formal Verification by 3.5x
TLA-Prover fine-tunes a 20B model with RL to generate verified TLA+ specs, hitting 30% pass@1 -- 3.5x better than any untuned LLM baseline
20 stories
TLA-Prover fine-tunes a 20B model with RL to generate verified TLA+ specs, hitting 30% pass@1 -- 3.5x better than any untuned LLM baseline
MDP-GRPO fixes three fundamental failure modes in GRPO training for instruction-following, boosting strict constraint satisfaction by up to 5% on Llama-3.2-3B
OPRD moves LLM distillation from output probabilities into hidden states, closing the student-teacher gap while training 1.44x faster and using 54% less GPU memory.
Yuanzhi He et al. propose iCEM+TL, a zero-shot transfer learning framework that boosts robotic manipulation success rates by up to 23% with no retraining required.
MemGate is a 9M-parameter plug-in that blocks unsafe memories from reaching your LLM agent without touching the model or the database
RBI-Eval is a new benchmark that measures when memory-augmented AI agents inappropriately surface sensitive user history in unrelated conversations
MRAgent replaces static memory retrieval with an iterative graph-traversal loop, cutting token costs by 5x while beating all baselines by up to 23% on long-conversation benchmarks.
A new metamorphic testing framework exposes when SHAP and LIME explanations are untrustworthy — without needing ground-truth labels
PLAN-S adds a style-conditioned cost map between a frozen world model and its planner, cutting collision rates by 42% and enabling controllable driving behavior without retraining
MGSD teaches vision-language models to plan spatially by distilling symbolic knowledge into visual reasoning, closing an 18-19% performance gap at inference time without any symbolic inputs.
RedditPersona is an open-source pipeline that turns raw Reddit data into community-conditioned LLM adapters, revealing a fundamental trade-off between how distinctly a model sounds and how natural it reads.
ATT-CR cuts transformer attention complexity from O(N²) to O(N) and filters cloudy pixels from attention, beating prior cloud removal methods on four benchmarks
CausalPhys introduces 3,000+ expert-annotated questions and a new fine-tuning method that teaches VLMs to reason about why things happen, not just what happens next
WLA-0 merges world modeling, language reasoning, and robot action into one 2B-parameter model that hits 92.94% success on RoboTwin2.0 at 40ms per inference
A new paper proves LLMs refuse to correct their own mistakes not because they can't, but because of how chat templates label the source of a claim — and a zero-training fix exists.
A new query-adaptive framework detects whether audio or video is actually useful per query before fusing them, beating fixed multimodal fusion by 4.2 points on a 12,000-video broadcast archive.
A new paper argues that AI in biomedicine must move from pattern recognition to forward simulation — and lays out the full blueprint for doing it
A new RL post-training framework teaches multimodal models to remember the full history of an editing session, not just the last instruction
UniVoice uses factorized conditioning and a learned null melody token to let a single 0.3B model match dedicated TTS quality while crushing prior unified singing baselines
RHO lets LLM agents rewrite their own tools and skills from past failures — no labeled data needed — lifting SWE-Bench Pro from 59% to 78% in one pass.