Aim the slop cannon: structure beats fluent wrongness
Agents, selective memory, and adversarial step-checks make LLM science inspectable
cs.AI
Artificial Intelligence
Covers all areas of AI except Vision, Robotics, Machine Learning, Multiagent Systems, and Computation and Language (Natural Language Processing), which have separate subject areas. In particular, includes Expert Systems, Theorem Proving (although this may overlap with Logic in Computer Science), Knowledge Representation, Planning, and Uncertainty in AI. Roughly includes material in ACM Subject Classes I.2.0, I.2.1, I.2.3, I.2.4, I.2.8, and I.2.11.
sort pith recommended most recent
Reward-based model tuning becomes provable Newton steps
Fine-tuning and sampling share one iteration with global descent and quadratic convergence.
· “Newton Matching for Generative Modeling: A Unified Framework for Fine-Tuning and Sampling”
Model agreement zeroes noise in private text generation
Each generated token leaks at most bT of mutual information about the private secret; unanimous votes add no noise.
· “PAC-Private Autoregressive Generation: Calibrating Noise to Ensemble Disagreement”
Two feedback tokens per answer can match or beat full training
Training a student model on one or two tokens per response matches or beats dense supervision in nine configurations.
· “Extremely Sparse Supervision Incentivizes Reasoning Ability”
Self-critique has a provable error floor in LLM reflection loops
Bilevel game theory shows only grounded verifiers converge to zero risk; the gated system resolves 72.2% on SWE-bench.
· “Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems”
Streaming ASR commits words 27–84 ms after they end
Wait when uncertain, emit when ready: latency drops from hundreds of milliseconds to tens without losing accuracy.
· “X2Streaming-ASR: wait when uncertain, emit when ready for streaming ASR”
QSP yields a solvable model of quantum representation learning
For sparse data the full nonlinear dynamics reduces to independent scalar flows with explicit convergence times.
Exact Mapping Links Frustrated and Field-Driven Spin Chains
A theorem uses Householder reflections to show a frustrated zero-field O(n) chain equals a simpler field-driven chain for any length and…
Independent learning in discounted games hits a quasi-polynomial barrier
Under ETH for PPAD, polynomial CCE computation is impossible; first radically uncoupled algorithm matches the hardness bound.
· “Independent Reinforcement Learning in Discounted Markov Games”
GPU-CFR ends the CPU advantage in tabular CFR with static dataflow and graph replay
A heads-up no-limit turn subgame with 83k info sets solves at 0.397ms per iteration, 14–258× faster than the fastest CPU solver.
Concept shift defined even when supports mismatch
Entropic optimal transport yields a well-defined, estimable shift; new bound unifies both shifts.
A 300M species specialist beats every 2–8B edge VLM tested
Even the specialist loses 18 points on field images, so blur hurts everyone, not just general models.
· “Can Edge-Deployable Vision-Language Models Identify Species?”
New model links generative AI answers to business impact
GMMM estimates the causal effect of GEO and GEM by converting generated-answer occurrences into marketing inputs.
Adaptive AI behavior emerges from persistence alone
A minimal experiment shows adaptive control arising without any task objective, reshaping alignment for persistent agents.
· “Artificial Id: Drive and Persistent Alignment in Agentic AI”
Topology recognition does not transfer to action
MindTopo benchmark tests 14 models on five topological properties; every model reasons better than it plans, revealing a fundamental gap.
· “MindTopo: Can Foundation Models Reason in Topological Space?”
Fine-tuned DeBERTa flags LLM hallucinations at 93.2% accuracy
A detector-guided preference-optimization step also cuts a generator's measured hallucinations by over half.
· “Domain-Specific Hallucination Detection in Large Language Models”
AssayLoop finds 27.7% of hits testing only 5% of a CRISPR library
A learned acquisition policy combined with LLM priors achieves 5.67-fold enrichment over random on held-out screens.
· “Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens”
Two closed-form solutions merge nuclear and Frobenius norm benefits
A unified framework shows that all state-of-the-art linear recommenders differ only in regularization type, enabling simpler scalable…
· “On the Regularization Landscape for the Linear Recommendation Models”
Five autonomy levels map the road to genuine self-improving AI
Headroom-Closed Index shows interactive tasks lag far behind knowledge; only L5 systems close the improvement loop on improvement itself.
· “The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement”
Self-correcting speech model gains 11 accuracy points on spoken math
Trained to verify every reasoning step, the model fixes its own errors mid-flight — at no extra latency.
· “RetroThinker: Enabling Retrospective Thinking in Speech LLMs”
A chat assistant parses users' model questions with 94% accuracy
It improves on the grammar-based baseline by 17 points and needs no fine-tuning for new model types.
Routing and content separate as LLMs retrieve facts
Layerwise interventions show a functional handoff where causal control shifts from query routing to answer content across depth
· “From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge”
Closed-form schedule cuts FID up to 38.6% in diffusion and flow models
Combining risk from a baseline checkpoint with kinetic action yields optimal time allocation; a frozen template recovers most gains.
· “Model-Aware Schedules Improve Generation via Fiberwise Optimal Transport”
Captains Want AI help—But Only If It's Reliable
Survey of maritime professionals reveals conditional acceptance and strong concerns about overreliance and skill loss.
· “Understanding Operator Attitudes Toward AI-Supported Decision Making in Maritime Operations”
Parallel token sampling causes VAR image glitches
A small autoregressive refiner restores spatial coherence and beats a model twice its size.
· “Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling”
Looped flows top looped reasoning with 59% on ARC-AGI-1
By training a stateful denoiser on temporally aligned denoising tasks, the model stabilizes recurrence and scales with inference steps.
Kimi-Audio ties top WER yet leads on code-switching
Switch-localized metrics reveal a system's true ability to recognize Yoruba at language boundaries, hidden by aggregate WER.
LLMs preserve facts but fail to reverse news framing
Controlled test across three model families shows framing reversal under 7% even when correctly detected.
· “Recognizing Is Not Reversing: A Controlled Inversion Test of Fact-Preserving News Framing”
Four-parameter gating beats single-signal designs in 33 of 36 tests
Multi-coefficient per-token KL mixing on TweetEval shows directional gains, but wide uncertainty across three seeds.
Internalizing risk policies boosts recall by 15 points
SIRF embeds content rules into LLM weights via 70M synthetic tokens, achieving 71% recall at 95% precision under <1s latency.
· “SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control”
Low-rank post-training slashes LLM output length by up to 40%
LOCUS selects a task-aware LoRA subspace to cut tokens without changing the alignment loss or accuracy.
· “LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation”
Organization theory lifts embodied AI teams by 64%
New framework ORCH outperforms four prior multi-agent methods across 25 wildfire missions and eight language models.
· “ORCH: Organizational Principles Enable Collective Intelligence in Embodied AI”
Model with one-phone steps tracks reference emotion 6x better
A single-layer neural CDE hit 0.53 correlation to reference emotion intensity; the tuned baseline scored 0.08.
· “Continuous-Time Acoustic Modelling with Neural Controlled Differential Equations”
Time-based readout shrinks analog SNN area by 20×
Voltage-to-time converter avoids bulky current-scaling circuits, enabling compact analog neurons for memristive inference
· “A Time-Based Readout for Vector-Matrix Multiplication in Fully Analog Memristive SNNs”
Reverse posterior anchor outperforms voting for LLM teams
Joint forward-reverse fusion consistently beats voting, electoral rules, and LLM judges on a medical diagnosis benchmark.
LLM-inferred priors cut CAD surface fitting error by 71%
Language-driven semantic reasoning guides geometric kernels toward design-intent surfaces without solver changes.
· “Language-Augmented Semantic Priors for B-Spline Surface Fitting”
Ray-scaling operator guarantees 100% safe robot actions
A differentiable safeguard aligns training and inference, preserving task success while enforcing hard constraints.
Bandit-guided evolution tops skill search at 55% lower cost
One scorer allocates costly executions and trims the weak, giving stronger agents on 50 examples per task.
· “COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization”
Train agent harnesses 1.84× faster by sorting recurring failures
Prioritizing cross-task harness flaws lifts reasoning accuracy 18.56% and aids transfer to new models.
· “Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents”
Graph connects 43,991 geospatial datasets across Harvard Dataverse
7,654 policy-relevant datasets reachable through place-based search, even with messy metadata
· “Geospatial AI, Dataverse Metadata, and the Study of Place-Based Government”
Logic is the study of inferential entitlement
Warrant theory redefines logical consequence and failure through the presence or absence of warrant, shifting focus from truth to…
Align AI by earned autonomy through staged norm learning
A developmental framework: agents internalise norms through embodied, motivated, staged interaction.
ZipCodec achieves 6.25 Hz streaming speech coding
At 0.80 kbps and 160 ms latency, it matches or exceeds non-streaming and streaming codecs in reconstruction and downstream tasks.
Diffusion model generates realistic smart-meter data under privacy constraints
Conditional generation of year-long sub-hourly load curves matches real data in forecasting and appliance detection, enabling safe energy…
· “LoaDiff: Conditional Generation of Electricity Consumption Time Series for Energy Analytics”
Memory agent holds 0.951 quality across 180 changing requests
Keeping its program, plans, and search state, it beats baselines that regenerate code each turn.
· “MAPLE: Memory-Augmented Planning with Language and Evolution”
Inverse models make modular-production RL 3× faster
By learning only state trajectories instead of raw actions, model-based RL cuts overflow and training cycles on a real testbed.
LLM blends alt data and earnings calls to beat consensus forecasts
Outperforms OLS, gradient-boosted trees, and analyst consensus across four commercial alternative data channels.
· “Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting”
Canonical region graph steadies CAD neural nets across file variations
Merging faces by underlying surface and using the solid's own coordinate frame eliminates catastrophic failures from repatitioning…
One method gets 0.419 exact match on SciQA with no fine-tuning
ShEx schemas + few-shot queries let LLMs write SPARQL; a sampling trick cuts metadata runtime 80×.
· “Enabling Knowledge Graph Understanding at Scale with the EXplore Your Graphs ENgine (EXYGEN)”
Imaging modality reshuffles CNN melanoma rankings
ResNet50 is most stable, but histopathology images reveal performance gaps invisible in dermoscopy alone.
· “A Comparative Evaluation of Pre-trained Convolutional Neural Networks for Melanoma Detection”
PFI-based power allocation recovers 63% of lost training throughput under caps
By measuring job-level power elasticity with a new metric, data centers can cut power without crippling model training.
· “Characterizing Job Power Elasticity for Power-Flexible AI Training”
Hidden prompt revision layer injects cultural stereotypes into AI images
Controlled experiment isolates the text-to-text preprocessing step as a causal source of bias, not just the image model.
· “Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems”
LiDAR cone detection hits 98% F1 in 3 ms on CPU alone
A Random Forest pipeline with seven geometric features outperforms GPU-dependent systems for Formula Student racing teams.
Joint diffusion produces better design templates than sequential generation
InterIL generates image and layout in one pass using a learnable communication module, beating prior methods on harmony and human…
SMT solvers learn variable clauses
Resolving on original non-ground clauses yields reusable learned clauses and potentially exponentially shorter proofs.
Non-language data lowers loss but not linguistic skills
Pre-pretraining on music, grammars, and cellular automata accelerates next-token prediction but fails to improve downstream benchmarks.
One-shot correctness score from a 96KiB activation map
Compressing the full hidden trajectory into a fixed tensor matches dense detectors with 67 times less storage and no extra LLM calls.
· “ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps”
Two-layer graph queries process docs at 95% accuracy
Dual graph architecture beats vector-RAG by 19% on free-form answer tasks in 38-report Sanofi test set.
Stale batch-normalization statistics move published unlearning numbers
Refitting only the statistics on retained data shifts 47 of 221 checkpoints and can reverse the pass/fail verdict.
Convention gap: humans succeed 26% beyond literal info
Exactly computable in Hanabi, the metric separates human and AI play and suggests AI compatibility with humans, not other AIs, matters most.
· “The Convention Gap: Towards Measuring Implicit Communication in Cooperative AI Evaluation”
Articulatory features make accent distances interpretable
Optimal transport on articulatory features matches black-box accuracy while identifying specific phonetic differences like rhoticity
RouteRepair cuts TSP optimality gap by 56.6%
The methodology also improves ACO priors over hand-designed baselines and transfers to larger CVRP instances.
LLM direct generation beats alignment for clinical annotation transfer
Strict F1 improves by 0.06–0.15 across six languages without task-specific training or candidate ranking.
· “Cross-Lingual Clinical Annotation Projection as Constrained Text Generation: A Six-Language Study”
Prevalence governs dataset precision
End-to-end measurement shows naive cross-pool precision transfer errs by 422%.
· “Prevalence Determines Precision:Silent Contamination in Detector-Defined Datasets”
Catastrophic forgetting is a head problem in sound classification
A study on FSD50K and AudioSet shows that complex feature-level continual learning methods are unnecessary for in-domain audio…
· “Investigating catastrophic forgetting in sound event classification”
Calibrated confidence unifies model cascades
A single shared threshold with provable reliability lets cheap models handle easy cases without retraining or per-model tuning.
· “Calibration-Aware Uncertainty Cascades for Efficient Heterogeneous Model Collaboration”
LLMs rank symbolic models better than they explain them
Clinicians favored comparative model rankings over term-by-term readings in a body-fat symbolic regression case.
Similarity windows boost multi-turn LLM routing by 16%
New method selects the best LLM per query while avoiding context loss and confusion, outperforming single-model baselines.
· “SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations”
Progressive pruning shrinks speech encoder by 20% with near-zero accuracy loss
X-AuT reduces audio-encoder depth from 18 to 14 layers while keeping word error rate within 0.14 points of baseline on ten public…
· “X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation”
Live deepfake callers flunk simple CAPTCHA-style tasks
Prompting callers to hum, sing, or puff cheeks spots impersonation far better than artifact scanning.
· “Deep-Fake CAPTCHA: Mitigating Next-Generation Social Engineering Attacks”
Live query narratives bridge SPARQL and domain experts
Data stories embed executable queries in prose, turning exploration of cultural-heritage knowledge graphs into quality assessment.
Stable confidence, not just high confidence, improves LLM reasoning
TASCO optimizes for confidence that stays robust under small perturbations, boosting accuracy by up to 17% while using fewer tokens.
· “Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning”
AI oversight curbs supplier environmental scandals
A multi‑country study finds that when buyers use AI to monitor environmental practices, suppliers have fewer media‑reported…
VikingRAG uses 5–32% of tokens to match top RAG accuracy
A directory-aware semantic storage and trace reuse system cuts token consumption dramatically while preserving high accuracy on structured…
· “VikingRAG: Accurate and Token-efficient Retrieval-augmented Generation over Structured Documents”
Interaction contracts keep agent behavior tied to user intent
A new software framework binds task-level goals to application effects through contracts that survive revisions, handovers, and policy…
· “Agent-Integrated Software: Interaction Contracts and Continuous Assurance”
Bluesky mod catches only 1 in 5 harmful posts
First independent analysis of transparent moderation system reveals 78% of harmful content escapes default service.
Sample-wise reliability weighting boosts EEG-EOG attention decoding
Outperforms unimodal baselines by up to 5.76% on AVGC and achieves 88.42% on DTU.
Emergent latent languages are private dialects
Six societies share semantics but not raw coordinates; an inherited global interface severely harms new learning.
One reified representation lets any GNN match dedicated KG models
Training on a single small knowledge graph for 30 minutes transfers to 40 unseen benchmarks.
· “Reification as a Transferable Vocabulary: Zero-Shot Link Prediction with Vanilla GNNs”
Cross-modal diffusion transformer boosts brain decoding by 4–7%
CoMA-DiT treats paired biosignals as mutual supervisors to improve attention and emotion decoding.
· “Exploring Diffusion Transformers for Cross-Modal Augmentation in Multimodal Brain State Decoding”
