HyperFeed

A new arXiv study evaluates large language models as numerical optimizers across convex, fractional, and constrained problems. The models often identify useful structure and produce near-optimal answers, but numerical errors, constraint violations, and formatting failures reduce reliability. Iterative feedback improves valid solutions and extends performance to larger dimensions, though persistent structural errors remain.

• arXiv paper 2610.37932 evaluates convex optimization, fractional programming, and constrained projection
• Models are compared against deterministic reference solvers across multiple problem dimensions
• Evaluation separates solution quality from numerical inaccuracies, constraint violations, and formatting errors
• Iterative refinement provides targeted feedback on feasibility and solution quality
• Performance depends strongly on problem structure and deteriorates as dimensions increase

SkillMem stores reusable skills as compact latent memories that guide agents during multi-step reasoning. A role-conditioned reader lets different agents use the same skill memory based on their roles and hidden states. The method distills behavior from textual skills while keeping the model backbone frozen. Tests on six benchmarks show gains for smaller Qwen3 models with low per-skill storage.

• 9.6 KB of storage covers each skill, including its retrieval embedding
• Eight by 256 memory configuration stays fixed across model scales
• Experiments use Qwen3 models ranging from 4B to 14B parameters
• Skills come from paired successful and failed multi-agent trajectories
• Six benchmarks evaluate SkillMem across multiple model sizes

GAUSS is a workload-aware simulator that predicts LLM serving latency distributions and sustainable request rates. It models prefill and decode separately, then uses queueing equations and Monte Carlo sampling to estimate time-to-first-token and time-between-token tails. Operators can test routing, traffic segments, hardware, and scheduling policies before deploying them. The simulator targets gaps left by single throughput figures and replay-based tools.

• Two outputs include full TTFT and TBT distributions, not only average throughput
• Lindley recurrence models stochastic arrivals and queueing under continuous batching
• Monte Carlo ensembles estimate latency tails that single traces undersample
• GAUSS evaluates hardware and workloads that have never been deployed
• Scheduling analysis includes shortest-remaining-processing-time capacity questions

A new study examines how healthcare professionals view algorithmic bias in machine learning and AI systems. Survey and interview results show that clinical and administrative staff can differ in how they judge causes, fairness, evaluation criteria, and mitigation methods. Those differences may shape whether professionals accept algorithm-generated outcomes. The authors call for more research on fairness in healthcare systems.

• arXiv paper 2610.38006 uses surveys and interviews with healthcare professionals
• Clinical and administrative roles produced different views on bias causes
• Participants assessed algorithm fairness and mitigation approaches differently
• Algorithm characteristics also influenced perceptions of fairness and acceptance

Researchers built a 250-question benchmark to test language models on inertial confinement fusion. The evaluation covered scientific reasoning, calculations, symbolic derivations, and equation writing, using automated checks and model judges. Prospero, specialized with domain pretraining and supervised fine-tuning, gained 10.7 percentage points over its base model. Frontier models still made serious numerical physics errors, so higher scores did not prove reliability.

• 250 questions span qualitative reasoning, numerical calculation, symbolic derivation, and equation formulation
• 10.7 percentage points marked Prospero’s gain over its base model
• Deterministic numerical and symbolic checks supplemented language-model judging
• Retrieval helped a smaller model but not a frontier model significantly
• Two stress tests exposed weaknesses in physical assumptions and model validity

PromptSense proposes a mental health chatbot framework that analyzes sentiment and depression risk before model inference. A rule-based filter then builds prompts tailored to the user’s emotional and safety context. The authors evaluated four model strategies and two prompt variants across empathy, helpfulness, clarity, relevance, safety, and toxicity. Prompt engineering beat DSFT by 1.8% and RLAIF by 1.1%.

• Four model strategies were tested alongside two prompt-engineering variants
• Evaluation covered empathy, helpfulness, clarity, relevance, and safety
• Perspective API supplied toxicity assessments for generated responses
• Rule-based filtering ran before model inference to identify emotional risk

A new study argues that strong LLM results on the CLadder causal reasoning benchmark may reflect pattern matching rather than genuine causal reasoning. The researchers found simple random forests also perform well on generated data, while commercial models remain brittle when key text is removed. The findings call for caution when interpreting benchmark gains from procedurally generated tests.

• CLadder was released in 2023 as a procedural causal reasoning benchmark
• Random forests performed well after training on generated CLadder examples
• Removing critical text sections had almost no effect on many LLM results
• Researchers attribute performance to patterns learned from generator data

A reinforcement learning agent improved warehouse product placement in a FlexSim digital twin simulation. The system learned from streaming operational data and adapted slotting decisions to reduce picking time and travel. Across a four-zone warehouse with forklifts, it also increased throughput, showing how agents can optimize physical operations through continuous feedback rather than fixed rules.

• 35.42% throughput increase reported across the simulated warehouse
• 18.54% shorter travel distance achieved with dynamic product placement
• Four-zone layout modeled forklift-based material handling operations
• Asset Administration Shell connected the agent, physical assets, and twin

ProACT is a virtual reality firefighter training system that adjusts fire behavior as trainees perform. Its global module changes scene-level fire spread in real time, while a local module updates object flammability periodically. Tests found better objective performance and user experience than adaptive and fixed-difficulty baselines. The approach could support other immersive training settings.

• arXiv:2610.38117v1 introduces ProACT for adaptive firefighter training
• Global control changes simulated fire spread during each training scenario
• Local control updates object-level flammability across the environment
• Evaluation compared ProACT with adaptive and non-adaptive baselines

Researchers introduce LLMON, a post-training method that separates executable instructions from passive data using explicit identifiers. Training Granite-4.1 models on 4.6 billion structured tokens sharply reduced indirect prompt-injection attack success without changing model architecture. The approach covers general, retrieval-augmented generation, and tool-calling data, while attention metrics expose layer-level failure modes.

• 4.6B tokens span general, RAG, and tool-calling training examples
• Granite-4.1 models trained at 3B and 8B parameter sizes
• Attack success fell from 53-100% to 0.02-5.6%
• Training used supervised fine-tuning followed by reinforcement learning
• Distractor variants added competing instructions and irrelevant context

A new paper argues that AI systems should decompose difficult tasks instead of expanding the model. Deterministic subtasks move to the host, leaving models to handle only the fuzzy remainder. Across nine models, this produced identical operation-selection scores and nearly identical selector-resolution scores, without fine-tuning. The authors report the pattern across physics, motion retargeting, object editing, and Web-CLI tools.

• Nine models span 494M parameters to a frontier Opus-class model
• Local checkpoints cover a 30-fold parameter range
• Operation selection scored identically across all nine models
• Selector resolution matched across eight of nine models
• Labelling remained the only task with a sloped performance curve

SkillCT introduces a benchmark for evaluating agent skills beyond whether they complete downstream tasks. It measures specification quality, behavioral robustness, semantic alignment, and responsibility boundaries across static, dynamic, and human-assisted dimensions. The benchmark also proposes an execution-free baseline that predicts quality profiles for unseen skills, enabling scalable pre-deployment checks without task-specific tools or repeated runs.

• Nine dimensions span static, dynamic, and human-assisted skill assessment
• Dynamic tests compare behavior across backbone models and request variations
• Human-assisted evaluation measures behavior-chain consistency and task granularity
• A public skill corpus provides reference data for quality weaknesses
• Execution-free prediction targets unseen skills without repeated task runs

LRMoE is a memory-aware training system that reduces sparse Mixture-of-Experts overhead on GPU-constrained nodes. It recalibrates backward demand, places experts under memory limits, and compacts retained token-expert pairs before communication and grouped matrix operations. Tests on Qwen3-30B-A3B and GLM4.5-106B-A12B reached fixed-loss targets 21.2% to 44.8% faster than SYMI.

• 21.2-44.8% faster cumulative MoE-layer training than SYMI
• Qwen3-30B-A3B used as one evaluation model
• GLM4.5-106B-A12B used as the larger evaluation model
• Backward selection preserves routed computations while reducing retained pairs
• Variable-size grouped GEMMs lower execution-buffer demand

WebAnything is an automated framework that reconstructs reusable website environments from real URLs for web-agent training and evaluation. Its Explorer-Builder captures browser-visible structure, resources, and interaction outcomes, while a verifier checks and repairs the results. The system produced 244 artifacts across 14 domains and supported 12,984 accepted tasks. Fine-tuning Qwen3.5-9B on its trajectories improved success across several benchmarks.

• 244 artifacts span 53 website categories and 14 application domains
• 12,984 executed tasks passed WebAnything’s acceptance checks
• Blind audit recorded 12 wins, two ties, and no losses versus WebHarbor
• Qwen3.5-9B training used 5,856 interaction trajectories
• WebArena-Verified Overall success rose 13.33 percentage points

A new study evaluates ontologies as checkers for LLM agents answering questions over knowledge graphs. It compares curated schemas, LLM-generated ontologies, and schemas mined from data across public and private graphs. Curated ontologies can catch errors, but they also flag correct answers and may contradict their own data. On UniProt, curated checks were only marginally more precise than random flags on real agent answers.

• Nearly 1 billion triples made up the study’s private knowledge graph
• UniProt’s curated ontology flagged more correct answers than two alternatives
• Tests covered open-weight and proprietary agents across multiple graph settings
• Checks sometimes missed structural errors even when ontology feedback was available

LibStructBench introduces a benchmark for testing whether AI agents can reconstruct sequencing-library structures and molecular workflows from protocols. Frontier agents often reach approximately correct answers but fail on exact molecular states and workflow connections. Changing the language model mattered more than changing the agent harness, while added knowledge and accumulated rules did not reliably improve held-out performance.

• LibStructBench targets customized sequencing assays and document-grounded protocol analysis
• Molecular-state reconstruction showed the only statistically significant harness or model comparison difference
• A small pilot improved oligo-family recovery but not mean workflow scores
• Accumulated procedural rules produced non-monotonic validation results across protocols

A new study finds socio-demographic information helps language models predict subjective offensiveness mainly at group and population levels, not for individual annotations. Across three multi-annotator datasets, demographic gains were small and inconsistent for single judgments but clearer after aggregation. Explicit reasoning improved results by separating a shared baseline from demographic-specific deviations, showing where the signal enters the prediction process.

• Three multi-annotator datasets tested prompted and trained language models
• Aggregation reduced idiosyncratic variation that hid systematic demographic signals
• Reasoning decomposed judgments into baseline and demographic-specific deviation
• Interventions linked better component estimates to improved final predictions

Researchers introduce Seq@K, an evaluation protocol for testing whether language-model agents solve tasks across K attempts with feedback after each failure. The method isolates iterative refinement rather than long-horizon planning, giving builders a clearer measure of how agents use feedback. Tests across seven benchmarks and 16 frontier models found different rankings and more efficient model configurations than standard Pass@K evaluations.

• 7 benchmarks tested Seq@K across 16 frontier language models
• K sequential attempts include feedback after every previous failure
• Model rankings often changed compared with Pass@1 and Pass@K
• Trajectory analysis measured issue resolution and answer rigidity
• Feedback type influenced how effectively models refined their answers

VerifiedLab is a multi-agent harness that verifies scientific claims field by field before releasing them for downstream use. Its AI scientists divide work among planning, evidence extraction, specialist drafting, and critical review. Tests on perovskite solar cells recovered known trends and predicted devices reported a year later. The system also attaches reliability measures to inherited literature.

• arXiv:2610.37923v1 introduces the VerifiedLab framework
• Four roles include a principal investigator, librarian, specialists, and critic
• Perovskite solar-cell experiments recovered known physical trends
• The system anticipated devices reported one year after its analysis

Researchers propose amortized localized posterior inference for calibrating models whose parameters vary across physical contexts. A conditional normalizing flow represents context-specific posteriors while accounting for observation and context uncertainty. Tests on the Branin-Hoo model and an ocean turbulence closure reduced aggregate prediction errors versus global or deterministic calibration. The method targets settings where one parameter cannot fit every context well.

• Branin-Hoo tests compared global and localized inference under uniform and skewed contexts
• Gaussian localization weights included covariance inflation for uncertain contexts
• Boundary reflection handled edge effects in the context domain
• Ocean tests used atmospheric surface forcing as the calibration context
• Temperature and zonal-velocity errors fell against deterministic and global calibration

ThinkJev uses one hybrid-thinking model to answer typed decisions quickly, then spend more tokens only when needed. It reads probabilities from the same option letters before or after reasoning, so escalation reuses the cached prompt and one sampled thought supports every budget. On 8,359 decisions, Qwen3.5-27B safely answered 24.5% directly and reached 81.6% accuracy with 72% of always-thinking tokens.

• 5% target produced 3.9% realized error for Qwen3.5-27B direct answers
• Qwen3.5-4B and 9B handled 6.5% and 0.0% directly
• Always thinking reached 82.7% accuracy on the same decision set
• Confidence stopping retained 81.6% accuracy while using 72% tokens

OpenFrontiers introduces a corpus of 12,151 open research problems and a 200-problem benchmark for scientific rediscovery. The benchmark uses Lean 4 statements, automated probes, independent reviews, and sandboxed proof checking. Seven LLM agents rarely rediscovered resolved problems without outside information, solving at most seven of 40. Live web access mainly enabled retrieval of published solutions rather than new discovery.

• 12,151 problems span multiple disciplines with recorded context and research progress
• 200 benchmark problems include 110 resolved in 2026 and 90 still open
• HQ-50 filters the benchmark by quality scores for focused evaluation
• Three settings compare closed-book, frozen-retrieval, and live-web performance
• Two independent checkers grade formal proofs inside an isolated sandbox

Researchers propose Partner-Outsider Positive Listening, a replay-based test for covert communication between AI agents. The method compares how a trained partner and an uninvolved replacement change actions after messages are reworded without changing their meaning. In simulated markets, PoPLi separated first-letter codes from ordinary communication, while message-only monitors missed or overflagged collusion. The test finds private channels, not collusion driven only by incentives or semantic meaning.

• PoPLi scores ranged from 0.48 to 0.59 whenever the covert code was present
• No-code conditions produced scores between -0.10 and -0.01
• All code-present 95% intervals stayed above 0.29
• A speaker-message monitor answered no in all 50 simulated markets
• A full-message monitor flagged 9 of 10 markets without collusion

Researchers evaluate how seven visual representations organize Indian foods beyond simple dish recognition. Using an expanded Khana dataset, they compare image retrieval, dish neighborhoods, culinary families, and recipe-based similarity. The results show that image-level neighborhoods differ widely and often cross culinary boundaries. Aggregated dish representations reveal shared visual structure, but it aligns weakly with recipe-derived relationships.

• Seven pretrained representations span convolutional, transformer, self-supervised, and image-text models
• Expanded Khana dataset supports image-to-image and image-to-dish retrieval tests
• Image neighborhoods sometimes connect foods from different culinary families
• Dish-level aggregation finds stronger visual agreement across individual images
• Visual similarity matches recipe-derived similarity less closely than expected

Verification Learning Theory proposes a process model for how AI assistance changes later unaided judgment. It links perceived assistant reliability to evidence checking, then links checking histories to retention of rule use. The theory predicts that expecting AI mistakes can improve delayed performance even when no errors appear. Simulations explore its assumptions, while experiments are still needed for validation.

• arXiv identifier: 2610.37875v1 marks the work as a new submission
• Strict VLT assumes equal learning from each genuine evidence check
• Perceived assistant fallibility may improve delayed judgment without observed errors
• The model separates retention of practised rules from choosing the right rule
• Simulations test parameter effects but do not provide empirical validation

FilingLoop introduces an interactive benchmark and self-evolving agent framework for reviewing and revising corporate disclosures. It reconstructs 380 IPO review tasks from SEC correspondence across 96 companies, testing whether agents can find missing facts, gather evidence, and satisfy revision requirements. Procedural feedback improves later cases without changing model weights, but overall requirement fulfillment remains below 30%.

• 380 IPO review tasks span 96 companies under distinct information constraints
• 6,099 eligible requirements support separate evaluation of filings and responses
• Requirement fulfillment rose from 24.28% to 27.96% after procedural evolution
• Full comment resolution increased from 16.90% to 20.75%
• Updates affect later cases only, avoiding retries on scored tasks

Researchers introduce HRF-Res, a spiking liquid state machine built from Harmonic Resonate-and-Fire neurons instead of standard LIF units. The resonant neurons respond selectively to input frequencies, helping the reservoir separate temporal patterns that low-pass dynamics blur together. Across ten benchmarks, HRF-Res reached state-of-the-art accuracy among spiking reservoirs while using up to 19 times less energy than an equivalent non-spiking oscillatory system.

• Ten benchmarks covered time series, sequential images, audio spikes, and neuromorphic vision
• Harmonic Resonate-and-Fire neurons replace Leaky Integrate-and-Fire neurons throughout the reservoir
• Up to 19x lower energy use than an equivalent non-spiking oscillatory reservoir
• Frequency-selective dynamics improved classification accuracy without removing event-driven spike computation

SciGame introduces a benchmark that tests whether AI agents can infer scientific laws through controlled experiments. It pairs historical and recent literature with original and altered environments, then measures candidate-guided identification and open-ended discovery. Results show familiar priors can mislead agents, while adaptive experimentation helps only stronger models. Exact symbolic law recovery remains the central failure point.

• Five frontier models were evaluated across identification and discovery tasks
• Temporal cards use papers published after models’ knowledge cutoffs
• Adaptive experimentation helped stronger models but hurt weaker ones
• The benchmark separates prior-only, fixed, and adaptive protocols
• Good explanations sometimes concealed failures on quantitative prediction