HyperFeed

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

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

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 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

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

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

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

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

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

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

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

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

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 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

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

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

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

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

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 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

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

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

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 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

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

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

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

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

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

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

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

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 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

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

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

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

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 risk-aware occupancy maps for safer end-to-end autonomous driving. The representation combines scene occupancy, traffic constraints, and future agent motion in a unified bird’s-eye-view map for trajectory planning. Their ROIDrive system reduces open-loop collisions by up to 52.9% on nuScenes, while RRDrive improves multimodal trajectory selection and candidate quality.

• 52.9% fewer collisions under UniAD, with 35.0% fewer under ST-P3
• ROIDrive injects occupancy predictions into planning queries for trajectory generation
• RRDrive reaches 0.951 selected-trajectory PDMS versus 0.937 for its baseline
• Challenging-scene candidate PDMS improves 30.2% with RRDrive
• RiskOcc4D-NAVSIM adds automatic risk annotations for trajectory evaluation

Researchers propose a framework that separates unmodeled cognitive effects from students’ affective states in educational response data. The model captures item, concept, student-concept, and student-item residuals before using affective signals to adjust guess and slip behavior. Its Q-matrix attention limits concept residuals to item-relevant concepts, reducing affect contamination. Tests span four datasets and six cognitive diagnosis backbones.

• Four datasets evaluated: ASSIST2017, ASSIST2012, ASSIST2009, and Junyi
• Six cognitive diagnosis backbones used for comparative experiments
• Residual components cover students, items, concepts, and student-item interactions
• Affective module adjusts guess and slip effects in responses

CogGym introduces a scalable framework for comparing AI models and humans on matched cognitive experiments. Its first release standardizes 258 experiments from 100 papers into a common format and tests 50 large language models. Larger, newer models track human judgments more closely, but progress on commonsense reasoning lags behind gains on math and coding benchmarks. Even the strongest models remain well below human consistency.

• 258 experiments drawn from 100 cognitive science papers in the initial release
• 50 large language models evaluated against matched human responses
• Human split-half reliability reached R² 0.93 for text, 0.95 for images
• Best models achieved only R² 0.59 on text-based tasks
• Experiment Markup Language standardizes diverse studies for reproducible comparisons

PlaceReasoner-Beta turns VLSI macro placement into a closed-loop, verifier-guided multi-agent task. A vision-language planner proposes layouts, while geometric and physical verifiers check legality and implementation results. A post-route optimizer then improves promising candidates using final power, performance, and area data. The accompanying PlaceReasoner-Bench evaluates methods on routed outcomes instead of proxy metrics.

• 16 tasks span 8 open RTL designs across two floorplan aspect ratios
• Geometric verification checks physical legality and expert placement rules
• Physical verification uses early implementation feedback to refine layouts
• Final evaluation measures routed PPA and design-rule violations
• Fixed floorplans and I/O assignments isolate macro positions and orientations

A production analytics agent study finds adaptive testing can cut recurring evaluations to 200 questions, or 38.5% of a full run, with 1.03 percentage points of error. The team instead deployed difficulty-stratified fixed subsets because they were easier to operate. Those subsets transferred to five other agent families without recalibration, offering a practical path for monitoring agents as they change.

• 574 historical benchmark runs supplied calibration and held-out evaluation data
• Multidimensional 2PL testing reached 1.03 pp MAE at 200 questions
• Fixed subsets stayed stable with calibration windows as short as one day
• Difficulty-stratified subsets transferred across five additional agent families

LEGIT introduces a protocol for helping buyers compare autonomous agents in emerging marketplaces. It binds measured quality and cost to a specific agent configuration, task domain, evaluation budget, and signed evidence. The system also links past outcomes to agent identities and models the deposits and fees needed for Sybil attacks. Its evaluations show that agent rankings can change with the testing budget.

• arXiv:2609.21325v1 presents LEGIT as a new credentialing protocol
• Signed records bind quality and cost to configurations, domains, and budgets
• Reputation records connect historical task outcomes to the same agent identity
• Sybil analysis estimates deposits and fees needed for reputation manipulation

Researchers propose Attention-Aware Routing, which uses recent attention patterns to guide expert selection in mixture-of-experts models. On OLMoE, the method improved GSM8K by 3.37 percentage points while leaving the base transformer frozen. The study finds routing changes can reshape attention in later layers, but applying the method too broadly may hurt factual retrieval. Layer-selective routing exposes a tradeoff between reasoning and retrieval.

• GSM8K improved by 3.37 percentage points over routing-only supervised fine-tuning
• OLMoE served as the test model with transformer weights kept entirely frozen
• Sliding-window features captured temporal and spectral patterns from attention weights
• Routing changes amplified attention sinks in the following transformer layer
• Deeper-layer intervention preserved math gains while indiscriminate use hurt factual retrieval

DENSE turns agent execution traces into reusable shortcut trees for self-refinement without post-hoc outcome labels. The method tracks local progress, recovery evidence, completed branches, and unresolved requirements across nested subtasks. On Terminal-Bench 2.1, it improved strict pass rates across four recipient models while reducing tokens used in reruns. The approach targets better agent retries with less expert supervision.

• Terminal-Bench 2.1 gains reached 7.12-15.64 percentage points over initial executions
• Reruns used 19.0-43.6% fewer observed recipient tokens
• REFIT resets environments and model contexts for fresh attempts
• DENSE ranked highest among tested non-privileged feedback methods
• GPT-5.5 ablations tested nested analysis, shortcut construction, and issue reconciliation

GVPO++ introduces Group Variance Policy Optimization, a method for improving LLM post-training without importance sampling. The approach derives gradient weights from the analytical solution to KL-constrained reward maximization, addressing instability in methods such as GRPO. GVPO also extends to on-policy distillation and supports a broader family of distillation objectives. Its main claim is a unique optimum with more flexible sampling distributions.

• arXiv:2609.21432v1 identifies importance sampling as a source of training instability
• GVPO optimizes the squared error between implicit and actual reward distances
• The method guarantees an exact solution to the KL-constrained reward objective
• GVPO supports flexible sampling distributions without importance sampling
• The framework extends to on-policy distillation and related objective families

LogicTrack is a neuro-symbolic framework that checks each large language model reasoning step with formal logic solvers. It scores logical soundness during inference and uses solver-based backtracking to revise flawed chains. The team also creates fine-tuning data from those traces, improving reasoning verification and final-answer accuracy across eight benchmarks and seven models.

• Eight reasoning benchmarks used to evaluate LogicTrack’s auditing framework
• Seven large language models tested in the experimental study
• Solver-Based Backtracking Reward scores each intermediate reasoning step
• Auto-formalization converts natural-language steps into symbolic representations
• Fine-tuning data includes backtracking traces from audited trajectories

PolyBridgeBench evaluates whether multimodal language models can design bridges that survive physics simulation, not just produce valid-looking structures. Models convert visual scenes and engineering constraints into node, member, and material topologies. The benchmark also tests whether they can repair failed designs using rollout evidence and a fixed interaction budget. Six models across 189 levels showed a wide gap between legal designs and dynamically successful ones.

• Six representative multimodal language models were tested across 189 bridge-design levels
• Deterministic legality checks gate every design before dynamic physics execution
• Failed rollouts return temporal visual evidence for the model’s repair attempt
• Material budgets caused pronounced changes in bridge-design performance
• Strict interaction budgets produced limited post-failure recovery

Researchers tested an offline multimodal language model for air-operations decision support in restricted environments. The system combines text and image retrieval from technical manuals, keeping answers traceable to doctrine without Internet access. In a pilot with four Brazilian Air Force image analysts, it matched human assessment scores while completing the task nearly four times faster. The study also measured the high mental demand of producing reports manually.

• 8/10 system score matched the human average on doctrinal knowledge assessment
• 7.1 minutes system completion time compared with 26.5 minutes for humans
• 6.0/7 mental-demand rating made manual reconnaissance reporting the hardest workload measure
• Four Brazilian Air Force image analysts participated in the pilot study
• Text and image inputs came from technical manuals in isolated environments

A new arXiv paper proposes Learning-to-Optimize as a missing layer in AI-native network design. Its framework uses offline optimization to generate supervisory knowledge, trains neural surrogate models, and applies them for low-latency runtime decisions. The approach shifts L2O beyond speeding up solvers toward a reusable architecture for autonomous resource management across communication and computing systems.

• Four stages define the workflow: optimization, knowledge generation, surrogate learning, and runtime inference
• NR-V2X relay selection serves as the paper’s example application
• Surrogate models inherit optimization expertise while reducing runtime decision latency
• The framework targets heterogeneous communication and computing systems
• The paper is listed as arXiv:2609.21519v1

RBS-Attention is a training-free method that makes long-context prefill faster by selecting only likely relevant attention blocks. Its centroid branch captures average relevance, while a radius-based rescue branch catches important tokens that averages can hide. On an H100 at 128K context, the method reached 5.97× faster end-to-end time to first token on Qwen3-30B-A3B-Instruct-2507-FP8. Quality stayed close to dense attention on RULER.

• 20.65× standalone prefill-attention speedup measured on H100 GPUs
• 11.92× vLLM prefill-attention speedup at 128K context
• 88.65 RULER accuracy versus 89.52 for dense attention
• LongBench-v2, InfiniteBench, and Video-MME supplied additional evaluations

DSRec introduces a sequential recommendation model that separates users’ long-term preferences from short-term session intent. It uses a full-sequence Mamba encoder for stable interests and a time-modulated state space model for behavior shaped by click intervals. A residual cross-fusion layer shares context between the branches while keeping their semantic roles distinct. The approach targets Transformer-level sequence modeling with lower computational cost.

• DSRec models item polysemy by assigning roles based on user context
• Short-term states adjust dynamically according to inter-click time intervals
• Long-term interests come from aggregation across the user’s history
• Residual cross-fusion aligns the two branches without merging their semantics

A round-trip study finds that language models lose substantial tree-structured information when converting expressions into word problems and back. Accuracy depends more on generation and extraction roles than on using the same model twice, with the best pairing reaching 92.9%. Most failures begin during generation, while tree depth and branching drive difficulty. Fine-tuning on roughly 3,600 examples improves every tested open-weight model.

• Sixteen models were evaluated in every generator-extractor pairing
• Role swaps changed round-trip accuracy by as much as 60.4 points
• Generation caused at least 73.6% of failed round trips
• Tree structure mattered more than model family for task difficulty
• Fine-tuning lifted open-weight models above untrained Gemini-3.1-Pro

Researchers introduce Embed-TTT, a two-step test-time training method for ARC-style visual reasoning. The method first trains a task embedding that represents the transformation rule, then freezes it while adapting the model backbone. Embeddings align more closely with known rules, improve task retrieval, and reveal compositional structure. Training fewer than 0.01% of parameters already solves some benchmark tasks.

• Three benchmarks include ARC-AGI-1, ConceptARC, and Mini-ARC evaluations
• Embed-TTT trains task embeddings before adapting the backbone
• Known-rule datasets support accurate linear probing of recovered rules
• The method recovers geometric structure in parametric transformations

CaLR is a new framework that uses causal structure and implicit differentiation to revise diffusion model reasoning steps. It treats reasoning as constrained latent optimization, allowing parallel generation to self-correct intermediate thoughts. The method targets a gap between autoregressive models, which can be locally greedy, and diffusion models, which often lack strict causal reasoning. Tests report state-of-the-art diffusion language model results and stronger Sudoku robustness.

• CaLR uses a causal topology matrix from an expert model
• Implicit differentiation guides revisions of intermediate reasoning states
• Parallel generation supports dynamic self-correction during inference
• Experiments report results against strong autoregressive baselines
• Sudoku serves as a constrained reasoning robustness test

Researchers adapted DINOv3 Vision Transformer models for automatic target recognition in synthetic aperture sonar imagery using LoRA. The method raised AUPRC from 0.300 to 0.679 while training only 0.26% of model weights. Hard-negative mining and supervised contrastive learning added no meaningful gains over matched controls. The results point to parameter-efficient adaptation as the main benefit for scarce, cluttered sonar data.

• Rank 4 LoRA reached 0.679 ± 0.027 AUPRC on at-sea SAS data
• Evaluation used a mission-level geographic split and three random seeds
• Comparisons were made at 85% test recall across all experiment arms
• Hard-negative mining changed AUPRC by −0.0045 ± 0.0119
• SupCon changed AUPRC by +0.0002 ± 0.0096

A new study detects LLM hallucinations by tracing the topology of information flow through attention graphs. Its single-pass method uses Forman-Ricci curvature and semi-local signals to identify attention patterns tied to incorrect responses. Across two benchmarks and multiple architectures, hallucinations correlated with weak context sharing, excessive self-attention, diffuse retrieval, and information squashing in the final layer.

• arXiv:2609.21096v1 introduces a single-pass hallucination detector
• Two benchmarks showed improvements over attention-based and multi-response baselines
• Forman-Ricci curvature identifies information bottlenecks in attention graphs
• Final transformer layers showed especially strong information over-squashing signals

A new study finds that fine-tuning reshapes large language model representations without consistently changing the components that causally drive task performance. Researchers compared attention patterns and layer activations with EAP-identified heads and logit-level features. Task-relevant components clustered in certain layers, but their locations rarely matched layers with the largest representational shifts. Shared components also failed to predict transfer and could coincide with negative transfer.

• arXiv:2609.21113v1 reports results on attention patterns and layer-wise activations
• EAP identified task-relevant attention heads and logit-level activations
• Component locations concentrated in specific layers, suggesting functional localisation
• Classification and generative tasks showed limited cross-task transfer despite overlap
• High component overlap could coincide with performance degradation on another task

TinyCeNN-LM presents a quality-gated method for replacing attention layers in pretrained language models with CeNN-inspired recurrent layers. The system accepts or rolls back each conversion using representation-fidelity and negative-log-likelihood checks. Tests accepted layers in SmolLM2-135M and Qwen3.5-0.8B, but results support cautious structural conversion rather than universal attention replacement or claimed speedups.

• Three variants were evaluated: Integrated Memory, MemoryFusion, and PDelta3-GDN2-CLVR+Local32
• SmolLM2 accepted layers 0 through 2 with cumulative ΔNLL of +0.01209
• Qwen3.5 accepted full-attention layers 3, 7, and 11
• Integrated Memory reduced total cache by up to 6.01%
• Converted Qwen releases scored 28.5% to 32.0% on 200 sampled items