HyperFeed

SPARE improves audio question answering by steering reasoning toward the final semantic conclusion before inference begins. The method adds a register token trained with Sentence-BERT similarity, helping models keep attention on audio instead of long reasoning traces. Tests with SALMONN show better zero-shot results on MMAU and MMAR, without increasing inference cost.

• MMAU and MMAR evaluations measured SPARE’s zero-shot audio reasoning gains
• SALMONN served as the backbone model in the experiments
• A cosine similarity loss aligned the register token with final conclusions
• Early attention to audio improved without adding inference-time computation

Bayesian Chronicle Agents introduce an explicit belief layer for LLM agents that separates internal stances from generated speech. Each stance is updated with one Bayesian step after an utterance, while a prior-strength parameter controls stubbornness. Sweeping that parameter produces consensus, persistent disagreement, or committed-minority influence. The design also makes agent behavior auditable and recovers prescribed stubbornness after language generation.

• Friedkin-Johnsen fixed points matched with R² values from 0.93 to 0.99
• One parameter, κ, controls agent stubbornness across simulation regimes
• Four language models showed perfect rank-order recovery of prescribed κ
• Explicit beliefs exposed stance biases that end-to-end simulations absorbed silently

PhysioBench introduces a unified benchmark for models answering questions about physiological signals. It combines 22 public datasets into 61.4 million questions covering 30 tasks, giving researchers one interface for varied clinical prediction goals. Tests of 21 model types found no system consistently strong across modalities and tasks. Results also show that question wording can materially change performance.

• 22 public datasets contribute annotations to the benchmark
• 61.4 million questions span 30 physiological signal tasks
• 21 model types were evaluated across three settings
• Each question links to a signal segment and source annotation
• Code is available on GitHub at Leanna97/PhysioBench

Researchers introduced μ²-Bench, a benchmark for testing whether multilingual language models truly remove harmful or private information. The benchmark follows the full cycle from memorization to unlearning and evaluation across multiple languages. It tests both languages seen during training and held-out languages, while measuring knowledge spread across languages. The results show that effective multilingual unlearning needs language-aware methods.

• arXiv:2609.20945v1 identifies the benchmark as a new research release
• Two evaluation settings cover both training languages and hold-out languages
• Cross-linguistic spread tests whether knowledge persists outside the target language
• The benchmark simulates memorization, unlearning, and evaluation as one pipeline

Researchers introduce ZENDAYA, a streaming schedule that controls how much video or audio a decoder reads before emitting text. Its single parameter links offline decoding and real-time generation, while estimating source consumption and latency in closed form. The method guarantees that outputs cannot use input that has not arrived, and experiments suggest reading less can improve text quality.

• Mean source consumption follows approximately 1/(1+γ) per emitted word
• ZENDAYA supports both ordinary offline decoding and real-time streaming
• The dependency guarantee holds for trained and untrained model weights
• The schedule extends to unbounded streams with arbitrary asynchronous arrival

Researchers propose Matrix Exponential Fixed-Point Iteration with Annealing, or MEFPIA, for finding equilibria in quantum games. The method works with local density matrices and tensor contractions instead of building the full joint quantum state. In tested games, MEFPIA matched Matrix Multiplicative Weights Update strategies and payoffs while reaching lower relative error in fewer iterations.

• MEFPIA uses effective Hamiltonians derived from payoff gradients
• The method targets extended Gutoski-Watrous games with local density matrices
• MMWU and MEFPIA reached the same tested strategy profiles
• Experiments measured convergence across multiple parameter settings

Researchers introduced a 12,000-instance multilingual benchmark for testing conflicting moral preferences in language models. GPT-5-mini favored Honesty over Autonomy across five languages without an explicit policy, while Llama-3.2-1/3B showed a strong first-option bias. Fine-tuning removed that bias with over 98% accuracy, and orthogonalized task vectors enabled models to switch ethical stances through task arithmetic.

• 12,000 dilemmas cover Honesty, Justice, and Autonomy conflicts
• Five languages include English, Hindi, Arabic, Spanish, and Chinese
• Llama-3.2-1/3B fine-tuning exceeded 98% accuracy
• Three value pairs test Honesty-Justice, Justice-Autonomy, and Autonomy-Honesty

Researchers propose a forced-alignment method that runs on end-user devices with linear memory instead of quadratic resource use. It combines Hirschberg’s algorithm with constrained random-walk pruning to align long speech recordings with text. Three-hour inputs use 5 MB rather than 140 GB and finish in one-third the CPU time of torchaudio. Pruning adds another 2x speedup on longer recordings while preserving accuracy in over 98% of tests.

• Hirschberg alignment matches torchaudio outputs while using linear memory
• Three-hour inputs require 5 MB instead of 140 GB
• CPU runtime reaches one-third of torchaudio on identical alignments
• Random-walk pruning adds 2x speedup beyond 20-minute inputs
• Pruning preserves alignment accuracy in more than 98% of tests

COAL-SQL combines coverage-guided augmentation with failure-driven learning to improve open-source models on complex text-to-SQL tasks. The framework finds missing SQL structures, generates complementary examples, and targets training at failures exposed during optimization. It keeps GRPO as the main objective while adding verified reasoning traces and related practice examples. The method reaches 64.9% execution accuracy using 12,600 post-training examples.

• 12,600 distinct examples support the complete post-training setup
• 64.9% execution accuracy reported on the evaluated text-to-SQL benchmark
• CGA uses greedy selection to find uncovered SQL structures
• FDL applies SFT to verified reasoning traces from a strong LLM
• Epoch-level retrieval creates practice examples tied to accumulated failures

Researchers propose DR-to-Long, a method that turns deep-research agent trajectories into long-context question-answering data. It replaces snippets and page summaries with full webpage contents while preserving evidence links across documents. The DLD-RL pipeline cycles through short deep-research reinforcement learning, zero-cost LongQA generation, long-context training, and another deep-research stage. The goal is to reduce errors from hallucination and missed cross-document evidence.

• 61.6% of remaining post-DR-RL errors came from insufficient long-context understanding
• DR-to-Long uses search histories, URLs, evidence snippets, and final answers
• Full webpage contents replace compact snippets without adding human annotations
• DLD-RL follows a DR-RL, LongQA-RL, then DR-RL training sequence

Proxifield is a decentralized protocol that builds sparse communication graphs between agents as their semantic relationships change. Agents route messages using direct address, information needs, plan alignment, and information complementarity, without training or a central planner. In tests, its advantage over a centralized Star protocol grew sharply with team size, suggesting better scaling for collaborative agent systems.

• 53.0% higher task reward than Star at team size N=25
• 59.5% higher task reward than Star at team size N=50
• 73.6% of no-failure reward retained after permanent agent failure
• Evaluated on Drone Search and Rescue and HiddenBench
• Performance improved from 35B to 397B parameter models

TALON is a radiology report generation framework that compares current scans with variable-length patient histories. Its Dual-Channel Temporal Fusion Module separates persistent findings from interval changes, then weights each prior examination by relevance. Tests on MIMIC-CXR beat the leading method across clinical and graph-based metrics. Performance also improves as more prior examinations become available.

• DCTFM uses separate similarity and change channels for prior exams
• Channel-specific attention estimates each examination’s relevance before fusion
• A learned gate suppresses redundant longitudinal evidence
• MIMIC-CXR experiments covered clinical efficacy and graph-based metrics

A three-stage ASR pipeline raises named-entity recall to 80-85% and filler recall to 76-86% for accented conversational English. It combines SQL-based data curation, regional LoRA adapters, and Qwen2.5-Omni-3B to produce verbatim and corrected transcripts together. The system targets language-learning feedback, where missed names and disfluencies matter more than WER alone. It also matches a zero-shot 30B model with one-tenth as many parameters.

• 2.8x entity density came from heuristic SQL filtering versus random sampling
• 80-85% entity recall rose from a 53-55% baseline
• 76-86% filler recall improved from below 5%
• 83.8% LLM-judge agreement covered 210 human-labelled samples
• 6k test utterances produced 6-10% WER across evaluated regions

TatBLiMP introduces the first benchmark for testing Tatar language-model grammaticality. It contains 1,248 minimal sentence pairs spanning 16 morphosyntactic phenomena, with each pair differing by one morpheme. Models are scored by whether they assign higher probability to the grammatical sentence, enabling evaluation without generation or parsing. The benchmark also tests whether focused Tatar training improves performance over broader multilingual exposure.

• 16 phenomena cover Tatar grammar, including number after numerals and quantifiers
• 1,248 sentence pairs use attested Tatar literary prose as grammatical examples
• Native speakers ratified every pair before inclusion in the benchmark
• Apertium-tat generates ungrammatical variants through deterministic morpheme changes
• MultiBLiMP’s 101 languages do not include Tatar

A scoping review of 48 studies maps how generative AI chatbots deliver motivational interviewing for health behavior change. Researchers found uneven safety reporting, favorable user perceptions, and limited evidence from real interventions. Most outcome studies covered only a single session, leaving questions about sustained behavior change, clinical effectiveness, and reliable deployment.

• 47 reports covering 48 studies met the review’s inclusion criteria
• 28 studies involved direct interaction with users, while 20 focused on system design
• 23 systems used dynamic adaptation to tailor chatbot responses
• 30 studies assessed motivational interviewing quality, often finding consistent interactions
• Only 18 studies reported intervention outcomes, mostly after one session

A new study finds that recursive language models can generalize better out of domain by isolating each subtask’s context. Standard chain-of-thought can use tokens outside the current subtask as shortcuts, fitting training data but failing when those tokens change. Recursive context limits block that failure mode, suggesting that reliable reasoning needs the right inductive bias, not just a broad hypothesis class.

• arXiv:2609.20831v1 presents the study as a new paper
• Standard CoT reads the full reasoning trace during learning
• Recursive models solve each subtask in an isolated context
• CoT can simulate the recursive rule with only a constant-factor cost

A study finds a frontier model achieves the best anxiety-detection F1 on Reddit, while a 110M-parameter encoder nearly matches it without an external API. The researchers show that benchmark scores are distorted because 69.3% of labeled posts contain “anxiety” or a variant. Removing those terms exposes large gaps in lexical dependence, especially among fine-tuned models.

• 0.846 F1 went to the frontier model on the original Reddit test set
• 110M-parameter encoder reached 0.831 F1 without external API access
• Mental-health pretraining added only 0.7 F1 points to the encoder
• 8.6 to 25.4 points measured the models’ lexical dependence
• LoRA fine-tuned 3B model relied on keywords more than TF-IDF

A new arXiv paper proposes evaluating human-AI collaboration by outcome quality divided by interaction cost, rather than task success alone. Across four tasks and two datasets, equally rated sessions differed by up to 70 times in interaction cost. The study finds task-specific tradeoffs, unreliable subjective productivity ratings, and better results when agents probe early and users do less repair work.

• 70 times separated sessions with identical quality ratings on interaction cost
• Four tasks across two datasets supported the proposed productivity framework
• Task demands determined whether extended interaction or fast convergence worked better
• Earlier agent probing correlated with more productive collaboration sessions
• Lower user repair effort marked sessions with stronger productivity

A new study evaluates how seven LLMs create and detect fake news across four controlled scenarios. Researchers generated 14,000 synthetic articles, then compared their language with real news and tested each model’s detection ability. Results varied widely by model, while the generation method strongly shaped detectability. Iteratively refined detection prompts often failed to help and sometimes reduced accuracy.

• 14,000 articles formed the synthetic fake-news corpus
• Four scenarios covered open-ended generation, rewriting, manipulation, and attribute prompts
• Models were tested with basic and iteratively refined detection prompts
• Generation strategy strongly influenced whether models could detect misinformation
• Refined prompts often harmed detection performance rather than improving it

HERMES is a graph-based framework that predicts patient outcomes using only clinical notes. It uses large-language-model-guided extraction to build personalized knowledge graphs with temporal links, treatment failures, and outcome changes. A Graph Attention Network then creates patient representations for mortality and readmission prediction. Tests on MIMIC-III and MIMIC-IV beat strong text-only baselines.

• MIMIC-III and MIMIC-IV evaluations covered in-hospital mortality and 30-day readmission
• Personalized graphs capture temporal dynamics and treatment failures from clinical text
• Graph Attention Network synthesizes patient representations from extracted clinical relationships
• Contrastive Logic Modeling distinguishes treatment changes and shifts in outcomes

CityLearn v3 is a configurable simulation and evaluation framework for testing renewable energy community controllers under realistic operating conditions. It models changing participants, equipment failures, service deadlines, demand-response requests, and electrical limits in one environment. The framework separates requested actions from actions actually applied, helping researchers detect infeasible commands and missed services that aggregate cost metrics can hide.

• CityLearn v3 models buildings, photovoltaics, batteries, electric vehicles, and flexible loads
• Declared timesteps preserve consistent power-to-energy accounting across simulations
• Reference controllers and service-aware indicators support comparisons across communities
• Synthetic high-frequency replay exposes short peaks hidden by annual energy totals

Ensemble QSP introduces a multi-agent framework that keeps long research workflows coherent across sessions with hierarchical memory. Its five specialist agents work under domain-expert principal investigators, while structured knowledge and physics checks enforce scientific constraints. The system achieved autonomous PKPD model selection and better parameter recovery than single-agent baselines. Tests with open-weight models suggest the design does not depend on proprietary models, though model capability still affects physical-consistency checks.

• 301-token median context injection, with a 4,050-token maximum
• Five specialist worker agents operate under domain-expert principal investigators
• PKPD benchmarks showed stronger parameter recovery than single-agent baselines
• DeepSeek-V4-Flash/Pro and Llama 3.1 70B replicated the findings
• Ablations separated memory, retrieval, and PI oversight failure modes

A new study finds that rewarding efficient reasoning helps language reasoning models recognize when they should abstain. Unlike humans, these models often spend more computation on unanswerable prompts, producing longer chains of thought. Fine-tuning several 4B models with a GRPO reward improved abstention by 12.8% on average while preserving answer quality and shortening reasoning traces by 44%.

• +12.8% average abstention gains matched human-like behavior on underspecified tasks
• 44% shorter chains of thought improved computational efficiency after fine-tuning
• Several 4B reasoning models received the new GRPO-based reward
• Human reasoning effort stayed bounded on unanswerable tasks
• Unanswerable prompts triggered longer model reasoning than answerable prompts

MME-Safety introduces a fine-grained benchmark for testing safety in multimodal large language models. Its four-dimensional annotations cover risk scenarios, harm severity, and modality-specific stealth levels. A hierarchical framework measures response reliability, risk exposure, and defensive behavior. Zero-shot tests across 17 leading MLLMs examine cross-modal attacks and the safety effects of chain-of-thought reasoning.

• 17 state-of-the-art MLLMs underwent the benchmark’s zero-shot safety evaluations
• Four annotation dimensions capture risks, severity, and modality-specific stealth
• Hierarchical scoring separates response reliability from actual risk exposure
• Evaluations compare multiple cross-modal input configurations and CoT reasoning effects

VISPATH introduces a multimodal knowledge graph question-answering framework that uses visual cues throughout multi-hop reasoning. Instead of relying on images only to find starting entities or retrieve evidence, it recomputes visual intent at each reasoning step. The framework also prunes full reasoning chains and checks whether selected evidence is sufficient before generating an answer. Its authors introduce VISPATH-Bench to evaluate the approach.

• VISPATH-Bench provides a new benchmark for multimodal KGQA evaluation
• Hop-specific intent combines the input, question, and current partial paths
• Candidate paths are scored as complete evidence chains before selection
• The framework checks evidence sufficiency before answer generation

BirdsongChat introduces a hybrid multi-agent system that converts multimodal instructions into interpretable controls for embodied simulation. LLM reasoning agents create a Unified Parameter Representation, while simulation agents generate synchronized motion, sound, and environmental behavior. Tested on interactive bird scenarios, the prototype targets better control and cross-modal consistency. Its explicit intermediate layer links language-level planning to physical execution.

• 94.4% cross-modal coherence score across simulated bird behaviors
• 100% affective consistency score for emotional behavior generation
• 92.6% generation consistency across tested scenarios
• UPR encodes behavioral states and interpretable simulation controls
• Tests covered text and image prompts, species, environments, and multi-bird interactions

SAGE turns grant rubrics into structured checks and links its judgments to evidence in applications. Tested on 35 nonprofit grant applications, it reached fair agreement with prior human reviews, then improved after foundation staff inspected and corrected its drafts. The system is designed to support expert review, not replace it, with claim-level audits marking confirmed, disputed, and missing evidence.

• 35 nonprofit grant applications formed SAGE’s evaluation set
• Kappa reached 0.58 in the assisted re-review
• One-prompt baseline scored kappa 0.33 on the common subset
• Foundation staff produced 202 criterion-level assessments
• Claim audits label evidence as confirmed, disputed, or unaddressed

Transsion’s Speech Team placed second in the MLC-SLM 2026 Challenge with a multilingual speaker-attributed transcription system. The cascaded pipeline combines speaker diarization, long-form ASR, timestamp alignment, and output fusion for conversational speech. It uses DiariZen for speaker segmentation and Qwen3-Omni for transcription. The system reached a 15.41% tcpMER on the official evaluation set.

• 15.41% tcpMER earned Transsion second place among all participating teams
• Three modules handle diarization, multilingual ASR, and speaker-transcription fusion
• Qwen3-Omni generates transcriptions across long-form multilingual conversations
• External CTC alignment supplies word- and character-level timestamps
• DiariZen creates speaker-homogeneous segments through clustering

Researchers introduce COPES, a dataset and evaluation framework for training language models to generate peer support grounded in lived experience. Fine-tuning improved strategy alignment by more than 50% for general-purpose models and also improved emotion and tone. Gains varied across subreddits and coping requests, while training shifted responses toward problem-focused advice and away from emotion-focused support.

• COPES evaluates support along three axes of community alignment
• Post-training used both supervised fine-tuning and direct preference optimization
• Alignment gains differed substantially across subreddits and coping strategies
• Training suppressed emotion-focused strategies in favor of problem-focused recommendations

Researchers propose a hypergradient-free method for steering self-interested agents toward socially optimal equilibria. The social-gradient flow uses the gradient of social cost rather than equilibrium sensitivities, which may be unavailable when planners lack agents’ cost functions. The paper proves convergence under observable equilibria and extends the result to two-timescale learning when agents’ actions must be inferred.

• arXiv:2604.11346v2 replaces the paper’s previous submission
• The method targets incentive design with unknown agent cost functions
• Agents update strategies faster than incentives in the two-timescale setup
• The proof covers any learning rule that asymptotically tracks equilibria

Researchers introduce progressive error curriculum training to make phoneme-to-text reconstruction more robust to realistic recognition mistakes. The method adapts an NLLB-based model with synthetic perturbations and pseudo-labels from visual speech recognizers. Tests on LRS2 and LRS3 show consistent gains across several visual speech recognition frontends. The work targets training and inference mismatch rather than autonomous agent behavior.

• LRS2 and LRS3 provide the evaluation benchmarks
• NLLB supplies the phoneme-to-text reconstruction backbone
• PECT combines synthetic errors with multi-domain pseudo-labels
• HP-VSR-FiLMFuse is among the evaluated visual recognition variants

A new paper describes loopjacking, where an agent executes operation B after a human approves operation A. The attacks exploit either misleading approval displays or mutable workflow state changed after approval. Researchers reproduced post-approval substitution across seven Agno AgentOS releases and 12 LangGraph Agent Server versions. Exact binding checks in OpenAI Agents SDK blocked the tested attacks.

• Seven Agno AgentOS releases through 3.0.9 showed post-approval substitution
• Twelve LangGraph Agent Server versions through 0.14.0 reproduced the attack
• OpenClaw 2026.2.23 exposed representation mismatch, fixed in 2026.2.24
• OpenAI Agents SDK versions 0.22.0 and 0.22.2 rejected mutated operations

A new study argues that the Voynich Manuscript is a structured pastiche rather than a readable natural language. Its models find letter-like symbols, Zipf-like word patterns, and phonetic features closer to Hebrew or Arabic. Image-text analysis also links its plant drawings to Pseudo-Apuleius herbals. Large language models support the image comparisons, but do not decode the manuscript.

• ArXiv paper number: 2609.20835v1, listed as a new submission
• Symbol analysis treats Voynich glyphs as letters, not syllabic units
• Phonetic patterns align more closely with Hebrew and Arabic than Indo-European languages
• Plant illustrations correspond strongly with Pseudo-Apuleius herbals
• Rare repeated initial-letter sequences point to deliberate language imitation

A study finds that token-level entropy cannot reliably measure uncertainty in small language models, but semantic entropy can. Across seven model pairs and five NLU benchmarks, semantic uncertainty helped route difficult queries to larger expert models, improving accuracy by up to 50 percentage points. Cross-family routing worked best, suggesting expert quality matters more than architectural compatibility. The payoff is smarter compute allocation, not lower compute use.

• 91% of dataset-model combinations showed near-zero mean token entropy
• Seven confidence approaches were tested across seven model pairs
• Five standard NLU benchmarks measured routing and accuracy gains
• Cross-family routing averaged 22.0% improvement versus 6.8% same-family
• Models under three billion parameters ran entirely on consumer hardware

CoLearn is an agentic tutor that tracks mastery and misconceptions, then generates each learner’s next question from that memory. Its persistent learner state uses soft-evidence Bayesian Knowledge Tracing, with a language model turning responses into continuous observations. In blind A/B tests, learners preferred memory-conditioned questions 68-69% of the time. The system also exposes its reasoning through live progress views and testing.

• 68-69% preference rate for memory-conditioned questions in blind A/B evaluation
• Three components combine learner memory, adaptive generation, and an evidence view
• Soft-evidence Bayesian Knowledge Tracing updates mastery as responses accumulate
• Persona simulations showed beliefs converging toward hidden ground-truth mastery

A new study examines how large language models can identify Kubernetes misconfigurations that threaten security and performance. The researchers propose a taxonomy of common errors and compare existing detection tools. They also examine which Kubernetes objects are most often misconfigured and assess issue severity. The work frames LLMs as a possible layer for improving cloud-native security checks.

• arXiv:2609.20834v1 introduces a taxonomy of Kubernetes misconfiguration types
• The evaluation benchmarks state-of-the-art tools for detecting configuration errors
• Researchers identify Kubernetes objects most prone to misconfiguration
• The study assesses the severity of detected security and performance issues

Reviser is a decoder-only Transformer that writes text through cursor-relative actions instead of fixed left-to-right decoding. It can move backward, insert tokens into earlier spans, and stop, enabling non-monotonic generation with one action per step. The authors report stronger arena preferences than SEDD and MDLM, competitive results at 100M and 300M parameters, and lower inference compute than multi-pass baselines.

• Three actions drive generation: INSERT(token), MOVE(Δ), and STOP
• 100M and 300M models match size-scaled autoregressive baselines
• Arena evaluations preferred Reviser over both SEDD and MDLM
• Trajectory data shows frequent backward moves and mid-canvas insertions
• Shared FLOPs comparisons found lower inference compute than diffusion baselines

Researchers introduced DischargeBench, a benchmark that tests whether LLMs can teach patients their discharge plans through open-ended dialogue. The simulation varies patient personality, education, health literacy, and medical recall, while a separate monitor agent preserves realism without changing the educator. Results show that aggregate scores hide failures across medical specialties and difficult patient personas. The benchmark prioritizes patient comprehension over polished text or answer accuracy alone.

• 477 cases span 24 ICD chapters in MIMIC-IV-Ext-DischargeBench
• Four scores measure conversation quality, coverage, comprehension, and factual consistency
• Virtual patients vary across personality, education, literacy, and medical-history recall
• Physician annotations align the LLM-as-a-Judge scoring process

A new study finds that subliminal prompting in language models depends more on causally active hidden states than fixed output geometry. Copying a temporary answer-position state between prompts shifted the model’s final animal score, while vector similarity and readout measures predicted behavior poorly. Tests across Llama and Qwen models also expose a multi-token length confound in digit scoring.

• +0.286 paired AUC change followed hidden-state copying across prompts
• Five intervention depths preserved the effect, including with eight blocks remaining
• All 18 tested concepts showed increased donor-control AUC
• Fixed output-vector correlation fell by 0.080 from Llama-3.1-8B to 70B
• Qwen digit scoring lost its association after controlling number width

A new arXiv paper proposes an agentic AI architecture to support Brazil’s Navy, Army, and Air Force. The design would let AI systems plan, access data, use tools, and act autonomously while remaining auditable. It targets faster observe-orient-decide-act cycles across land, aerospace, naval, cyber, and electromagnetic operations. The proposal also emphasizes permissions, security, and human oversight.

• Four use cases cover decision support, situational analysis, feasibility studies, and countermeasure suggestions
• Five operational domains include land, aerospace, naval, cyber, and electromagnetic spectrum
• Three Brazilian military branches are included: Navy, Army, and Air Force
• Safeguards address data access, permissions, security, and auditable autonomous action

A new study finds that LLM agent pipelines often drop the provenance showing whether an action was verified. When monitors saw the same authorization claim without its unverified framing, risky-action approval rose sharply across open and hosted models. Summarizers and memory compressors caused similar failures in full pipelines. The authors recommend carrying authorization provenance as structured state throughout agent workflows.

• Approval rose from 5% to 60% on Llama-3.1-8B after provenance removal
• Qwen2.5-14B approval increased from 9% to 98% under the same change
• Full pipelines produced 57% to 81% risky approvals across three monitors
• Tests used nine open-weight monitors and two hosted models
• Explicit rejection prompts sometimes blocked legitimate requests

ProKDA introduces a three-stage training method for explainable hateful meme detection. The approach first builds background knowledge, then learns hatefulness detection, and finally aligns decision boundaries. Its agentic knowledge pipeline supports the training process, but the paper focuses on multimodal model performance rather than deploying an autonomous agent. Tests cover three public hateful meme benchmarks.

• Three stages cover knowledge learning, detection learning, and boundary alignment
• External knowledge comes from an agentic background knowledge construction pipeline
• Joint explain-and-detect training can underperform simple SFT baselines
• Experiments evaluate ProKDA on three public hateful meme benchmarks

CliniCIRCA is a multi-stage LLM framework that reconstructs chronological mental health patient journeys from untimestamped discharge summaries. It extracts clinical events, assigns calendar-aware time labels, and produces date-grouped summaries for clinician review. The team built a verified benchmark from 52 summaries and scaled the approach to 1,000 silver-standard timelines. Instruction tuning improved five open-weight models across extraction, temporal tagging, and summarization.

• 14,882 MIMIC-III mental health admissions supplied the source records
• 52 discharge summaries produced 15,891 temporally tagged clinical events
• Clinicians corrected 629 errors during the evaluation process
• Generated timelines compress each source by a factor of 1.52
• Five open-weight models improved with instruction tuning

Researchers propose CC-OPI, an online task-allocation algorithm for UAV swarms operating with short communication ranges. The method interleaves task negotiation with physical execution instead of waiting for global consensus. Event-driven replanning, non-preemptive action locks, and fault-tolerant state synchronization help prevent stale assignments and endless reassignment. Simulations show stronger completion rates than matched PI and CBBA baselines.

• 250-meter communication radius used in the reported simulations
• Task completion rate reached about 0.80 under fragmented swarm connectivity
• CC-OPI improved matched PI and CBBA execution by roughly seven percentage points
• Version-based synchronization pairs with a global-time emergency pool
• Algorithm replans only after discrete physical or network-topology events

A new study finds full-duplex speech models speak mainly when addressed or after silence, not when content demands intervention. Across five model families, false claims and hazards rarely triggered spontaneous responses, even when pauses or permission to interrupt were available. Moshi and PersonaPlex answered most direct questions, but challenged false claims in only 14% to 15% of non-empty replies and warned about hazards in 4% to 7%.

• Five model families were tested across 10 turn-allocation conditions
• Moshi and PersonaPlex showed lower response probabilities after false-fact triggers
• Compressed pauses limited opportunities created by silence during testing
• Direct-question response rates were high when models received the floor
• Hazard warnings appeared in just 4% to 7% of non-empty replies

Neo-Classic tests whether language models can reason about Classical Chinese poetry beyond memorized historical patterns. Its contemporary expert-written poems create an out-of-sample challenge, and models lose 20 to 50 percent accuracy compared with historical texts. Discourse-ordering accuracy stays between 0 and 13 percent, showing that models handle local form better than global structure. Expert guidance raises reasoning-model performance to 36 percent, still below humans.

• Five behavioral probes test hierarchical constraint satisfaction in poetry reasoning
• Contemporary expert-authored poems reduce the chance of direct retrieval
• Qwen3-Max, Gemini-3-Pro, and DeepSeek-V3.2 were evaluated
• Reasoning-enhanced models reached 36 percent with expert-level guidance
• Discourse-level ordering accuracy ranged from 0 to 13 percent

A study found that language-model groups overstated consensus when replaying 100 human Wason reasoning groups. Agents participated more consistently, and their consensus rates stayed 34 to 44 percentage points above matched human groups after researchers adjusted for participation and scoring. Reasoning-mode agents often agreed nearly unanimously on incorrect answers, showing that simulated consensus did not predict collective accuracy.

• 100 held-out human groups were replayed with matched LLM agent groups
• Human full-consensus estimates ranged from 24.0% to 57.0%
• Submit-based comparisons found gaps of 34.0 and 43.9 points
• Participation-matched comparisons used 45 groups and gaps up to 44.4 points
• Reasoning-mode agents often agreed nearly unanimously on wrong answers

Reflective Recovery is a self-supervised method that turns failed reasoning attempts into training data for error recovery. It feeds partial, likely flawed trajectories back to the model alongside prompts, teaching it to recognize and correct mistakes without critics or reward models. On DeepSeek-R1-Distill-Qwen-7B, the method improved results on AIME 2025 and Minerva, addressing the limits of adding more perfect examples.

• AIME 2025 accuracy rose from 30.0% to 37.5% on the 7B model
• Minerva accuracy increased from 37.6% to 47.8% with Reflective Recovery
• Training uses failed trajectory segments instead of external critics or reward models
• The method targets Scaling Collapse from adding more positive reasoning examples

A new study examines cross-lingual transfer for sequential sentence classification across 13 non-English languages. It finds that linguistic proximity does not consistently predict performance, while similarity in rhetorical structure has a weak but reliable effect. Label-distribution similarity becomes the strongest predictor after accounting for source-language results. The authors also propose three structure-aware methods for generative models.

• 13 non-English languages drawn from five academic databases
• Encoder-based and generative models tested in transfer experiments
• Label-distribution similarity was the most consistent performance predictor
• Three proposed methods use structural information with generative models

Researchers propose a frontend-backend design that lets full-duplex speech models delegate tool calls to a text-based backend. A speech-to-text frontend emits a delegation token, while backend results return through a lightweight prefill-and-repeat step before streaming speech synthesis. The system preserves interruption handling and low latency, reaching 92-97% tool-call recall and beating GPT-realtime-mini and Qwen3-Omni-30B-A3B-Instruct on EVA-Bench.

• 81.2% accuracy rejecting irrelevant tool calls in single-turn evaluation
• Qwen3-235B-A22B backend delivered competitive Full-Duplex-Bench-V3 results
• Streaming ASR transcripts feed the backend LLM for tool selection
• Backend results re-enter the frontend through prefill-and-repeat injection

Researchers introduced ReaLMem, a benchmark built from authentic multi-year personal visual archives and first-person annotations. It tests AI systems on factual recall, persona inference, and predictive personalization rather than simple memory lookup. The accompanying ChronoProfiler weights attributes by temporal stability, helping systems resolve changing preferences. Results show predictive personalization remains a major weakness for current multimodal models and memory systems.

• Three evaluation tiers cover recall, persona inference, and predictive personalization
• Authentic multi-year visual archives replace synthetic, text-only memory datasets
• ChronoProfiler assigns temporal stability scores to user attributes
• Current frontier multimodal models show a ceiling on predictive personalization

FARSIGHT evaluated 15 academic financial trading-agent schemes for robustness and security under market stress. The study found 80% failed at least one core robustness metric, while every scheme showed security vulnerabilities. Because these agents can execute trades with real capital, small errors or cheap attacks may amplify into market-wide crashes. The framework tests both turbulence and adversarial behavior at the scheme level.

• 15 academic financial LLM trading schemes assessed by FARSIGHT
• 80% failed at least one core robustness metric
• 100% showed vulnerabilities across the security evaluation
• Three attack classes covered information sources, agents, and agent-as-attacker behavior

A new study analyzes 17,012 app store reviews across six major generative AI applications to measure trust, sentiment, and adoption barriers. Complaints cluster around advertising, authentication, server reliability, and subscription pricing. Claude shows the highest negative sentiment at 47.7%, but also a strongly enthusiastic user base, suggesting sharp polarization. The researchers propose a Trust Friction Score for comparing usability and trust problems across apps.

• 91% of advertising-related reviews carried negative sentiment
• Authentication complaints were negative in 89% of reviewed posts
• Server reliability drew negative sentiment in 83% of mentions
• Subscription pricing complaints were negative in 73% of mentions
• Researchers validated models against human coding of 300 reviews

Researchers introduce Goal-oriented Irregular Repetition Slotted ALOHA, a random-access scheme for massive sensor networks. GO-IRSA combines irregular repetition with belief-based transmission policies to prioritize task-relevant updates. The method avoids heavy sensor-side computation and frequent feedback while improving estimates of a distributed Wiener process. Tests report over 30% lower average and worst-case error than an optimal centralized solution.

• Over 30% lower average and worst-case estimation error than centralized access
• Thousands of sensors supported in the reported network experiments
• Distributed Wiener process used as the estimation task
• Robust to imperfect interference cancellation and inaccurate process knowledge
• Sensors avoid significant computing loads and frequent feedback

A new paper proposes MaxNorm-AC to stabilize advantage scaling in group-relative optimization for verifier-based reinforcement learning. The method separates meaningless reward jitter from small but credible gaps, then bounds how strongly those gaps affect updates. It targets failures where RLOO and Dr.GRPO underweight small differences, while GRPO can amplify them without limit. Tests across dense and MoE models cover math and code reasoning.

• arXiv:2609.19164 introduces the Reward-Resolution Protocol for filtering sub-resolution gaps
• Three-way calibration links reward strength, prompt weighting, and effective KL calibration
• RLOO and Dr.GRPO can make credible small gaps KL dominated
• GRPO’s standard-deviation denominator can amplify tiny gaps without bound
• Experiments span dense and MoE architectures on math and code reasoning

VisKG-LM compiles retrieved knowledge-graph subgraphs into images that language models can reuse as read-only visual memory. The approach moves graph encoding offline, so inference avoids rebuilding the same structure for every question and candidate. It improves several multiple-choice QA benchmarks while using about 400 million online parameters. The work targets more efficient graph-augmented reasoning rather than autonomous agents.

• 1.2 points better than GreaseLM on CommonsenseQA test sets
• 0.8-point gain over GreaseLM on OpenBookQA evaluations
• 4.3-point improvement over GreaseLM on MedQA-USMLE
• About 400M online parameters, versus GraphVis at 7B
• Relation-Labeled Paths preserve branching structure in rendered images

Researchers propose COIN-GP, a distributed framework that estimates system states and unknown dynamics from partial measurements. The method combines observer-based state estimation with online distributed Gaussian Process regression across sensor-equipped networks. A data collection strategy supports feasible learning, while theoretical bounds cover both state and model errors. Simulations outperform existing distributed GP methods.

• Paper number: arXiv:2609.20598v1
• Core method: online distributed Gaussian Process regression
• Setting: networked systems with partial state observations
• Guarantee: deterministic upper bounds for state and model errors
• Evaluation: simulations compare against distributed GP-based methods

VLM-CAD introduces a collaborative agent workflow that uses vision-language models to size analog circuits from schematics. Image2Net converts circuit images into topology graphs and JSON, giving agents structured facts instead of raw pixels alone. ExTuRBO then uses agent-generated seeds and sensitivity analysis to guide and explain Bayesian optimization. Across 12 tasks, the system reached 23.3% Strict Pass@1 and 91.7% Relaxed Pass@1.

• 12 sizing tasks covered six circuits across four technology platforms
• 23.3% pooled Strict Pass@1 measured fully compliant circuit designs
• 91.7% Relaxed Pass@1 showed broader near-compliant design success
• Automatic Relevance Determination supplied sensitivity evidence for reports

A new arXiv study evaluates 20 LLMs against China’s AI-generated content compliance requirements. The researchers test 2,303 questions across six dimensions, including 203 constitutional questions, and use multiple judges to assess compliance and refusal behavior. International models also comply at high rates, while the largest gaps appear on topics tied to ideological alignment. The benchmark gives developers a shared, legally grounded test for Chinese and non-Chinese models.

• 2,303 questions span six compliance dimensions in the evaluation framework
• 203 self-constructed constitutional questions test politically sensitive content behavior
• Several independent judges produce verdicts using hierarchical alignment memory
• International models show high compliance rates on standard Chinese prompts
• Ideological-alignment dimensions create the main differences between evaluated models

Researchers propose an offline self-distillation method that trains language models to follow deployment-specific safety boundaries. On political persuasion prompts, the method raised refusal from 9.47% to 84.75% with Qwen3-8B, while cutting unsafe responses across three benchmarks to 0.14%. However, over-refusal on XSTest reached 74%, showing that stronger safety can reduce useful answers without carefully balanced data.

• Single-shot generation left 19.88% of prompts without accepted refusal traces
• Escalating retries reduced missing refusal traces to 0.20%
• Verified target-model responses cut over-refusal from 15.20% to 5.20%
• Boundary-pair data lowered held-out comply-side over-refusal from 32.94% to 4.16%
• Harmful-side refusal fell modestly from 91.88% to 87.72%