❌

Normal view

Design Guidelines for Online Health Forums: User-Centered Design Approach

Background: Online health forums are used widely, yet evidence of their effectiveness is inconsistent. Evidence-based forum design guidance grounded in theory and lived experience could improve the efficacy and outcomes of these forums for the many people using them worldwide. Objective: This study aimed to draw on the experience of online forum users and staff, and insights from existing research on technology design and self-determination theory, to generate a set of theoretically grounded guidelines for safe and well-being–supportive online forums. Methods: We conducted 54 semistructured interviews (36 forum users and 18 forum staff) and 4 design workshops with forum staff, combined with input from a multidisciplinary research team. Principles of qualitative framework analysis were used to adapt a preexisting framework for well-being–supportive technology to the context of online health forums. Results: The resulting design guidelines are framed around 4 overarching principles relating to the psychological needs for autonomy, competence, and relatedness as defined by self-determination theory, and the additional need for safety in online forums. Each principle is presented alongside pragmatic design heuristics and specific implementation strategies. Conclusions: User experiences of online health forums are mixed. We have drawn on self-determination theory to propose evidence-informed guidelines for service development and refinement, adaptable to specific user groups across diverse settings. International Registered Report Identifier (IRRID): RR2-https://doi.org/10.1136/bmjopen-2023-075142

Evaluation of the Square Eyes Model as a Screening Tool for Identifying Digital Technologies in Wearable Camera Images Among Children: Laboratory Study

Background: Accurate measurements of children’s digital technology use are essential for understanding its potential implications on health and well-being. Wearable cameras can provide such measurements, but image coding is a high burden for researchers. Machine learning–based object-recognition models have the potential to reduce this burden by identifying images containing technology. Objective: This study aims to evaluate the performance of an object recognition model, the Square Eyes model, as a screening tool for identifying technologies in wearable camera images among children for further human review, as well as to examine the potential influence of face-blurring methods on the model’s performance. Methods: This study used data collected on 48 children (aged 3‐14 y) during an approximately 1-hour laboratory session. The children performed various technology-related tasks while wearing a camera. A total of 221,226 images were coded by humans and processed through the Square Eyes model. The performance of the Square Eyes model as a screening tool was evaluated by (1) assessing agreement between the model and human coding; (2) evaluating the N-back algorithm, an algorithm embedded in the model aimed to flag images requiring human review; and (3) examining the potential influence of facial-blurring on model performance. Results: Humans detected technology in 92,745 (41.9%) images, and the Square Eyes model detected technologies with an overall accuracy of 78.0%. When considering specific technologies, agreement between the model and human coders was the highest for (n=19,148, 54.3%) and (n=8492, 44.5%) and lowest for smaller devices such as (n=2600, 31.3%) and (n=3685, 25.1%). The model’s N-back algorithm effectively flagged images that required further human review, with only 7144 (3.2%) images that were not flagged for screening containing a human-coded technology. An explorative analysis indicated that using a square face-blurring with border could have reduced the model’s ability to accurately detect technologies. Conclusions: The Square Eyes model demonstrated overall satisfying accuracy in detecting technologies and successfully flagged images that required further review by humans. These findings suggest that the model could be used as an effective screening tool for reducing the burden of human coding. However, the model could be improved to more accurately detect smaller devices, and the form of facial blurring in images should be considered.
  • ✇MIT Technology Review
  • Donated livers can be made biologically younger Jessica Hamzelou
    Once an organ is removed from a donor’s body, the clock starts ticking. Surgeons usually flush the organ with a preservative solution, bag it, and put it on ice—where it immediately starts to degrade. The team has a matter of hours to get it into a recipient’s body. There’s another option—one that has been growing in popularity in recent years, especially for donated organs that aren’t in the healthiest state. Some hospitals opt to put them on machines that pump them with nutrients and rem
     

Donated livers can be made biologically younger

15 September 2026 at 00:11

Once an organ is removed from a donor’s body, the clock starts ticking. Surgeons usually flush the organ with a preservative solution, bag it, and put it on ice—where it immediately starts to degrade. The team has a matter of hours to get it into a recipient’s body.

There’s another option—one that has been growing in popularity in recent years, especially for donated organs that aren’t in the healthiest state. Some hospitals opt to put them on machines that pump them with nutrients and remove waste products, usually for around six to 12 hours. It’s a bit like being back in a body.

This allows doctors to assess the organs, and some recent studies suggest that time spent on these perfusion machines helps them do better once they’re transplanted. Now, scientists have found that perfused organs seem to get younger, at least at a molecular level.

The research, shared with MIT Technology Review, provides molecular clues as to why organs from younger donors are known to have a higher success rate. It might also help explain why perfused organs are less likely to fail once they make it into a recipient. 

The researchers behind the study hope to find new ways to test the health of donated organs and potentially develop additional tools to repair organs that might otherwise be discarded. “If [we] can improve the utilization of organs beyond what the current systems can do, then that’s a win in my book,” says Jesse Poganik, who studies aging at Brigham and Women’s Hospital in Boston and coauthored the study.

Clocking organs

Poganik—along with colleagues including Heidi Yeh and Alban Longchamp, transplant surgeons at Mass General Brigham—used “aging clocks” to assess donated livers. These are scientific tools designed to measure biological age—a result that is meant to convey more about the health status of an organ (or person) than chronological age.

In an initial experiment, the team used a clock to look at the patterns of chemical marks on DNA in 37 samples taken from 19 donated livers. Such epigenetic patterns are known to change as we age. But when the team compared samples from livers kept on ice and those that were perfused, the team found a “striking” pattern in the latter.

“Machine-perfused livers, in spite of being older or having other disadvantageous characteristics, had a biological age that was lower than [non-perfused] livers that were chronologically younger,” says Yeh, who led the work.

To investigate further, Yeh and her colleagues analyzed another 208 samples from 103 donated livers. This time, they used different aging clocks—ones that essentially measure how genes are working. They studied samples biopsied from the livers after they had been stored for up to around six hours either in cold storage or on machine perfusion.

In most cases, they also assessed a second sample taken around an hour after the livers had been transplanted into a recipient. Once the organ’s blood supply is reestablished in the body, “you have a few other things to do,” says Longchamp. “Then you just do a quick biopsy before you close.”

According to the clocks, which were developed to measure age and risk of death, the machine-perfused livers were biologically younger, the team found. “Pumping them at 34 degrees with oxygen and nutrients actually reversed the biological age,” says Longchamp. The results have been been shared with colleagues at an industry conference, he says. 

“If you adjust out chronological age … to have a fair head-to-head comparison, the difference between the two is on the order of 30%,” says Poganik. “It’s logical to say that perfusion drives this effect.”

The biological ages of all the livers tended to increase as soon as they were put into a recipient’s body, probably as a result of stresses on the organs. But still, the effect endured—the perfused organs remained biologically younger. 

Nathanael Raschzok, a transplant surgeon at Charité Universitätsmedizin Berlin in Germany who was not involved in the research, says the work is impressive. But it’s not yet clear what these changes might mean for the recipients of these organs, he says. The organs in the study were donated by people in their 30s, 40s, and 50s. Raschzok wants to know the effect of perfusion on the liver of an 80-year-old. “Every so often, we use organs from 70-, 80-, 85-year-old donors,” he says.

A better understanding of why the organs appear to be getting biologically younger might lead to therapies that achieve the same effect with a drug that could potentially be used to treat a donated organ for a fraction of the price, he adds. That’s important because perfusion is expensive—Raschzok says it costs around €10,000 in Germany (a quarter of the budget for a transplant), while the cost in the US comes to around $80,000 to $100,000 per organ, says Yeh.

Molecular repair

Yeh and her colleagues weren’t able to study most of the livers before perfusion. That’s because donated organs are generally not considered to be under the purview of the hospital until they’ve been placed on perfusion machines, she says. (Organ procurement procedures vary, but for the team as Mass General Brigham, donated organs are put on perfusion devices at the donor’s hospital. “There’s this sort of nebulous period where it’s not clear who the organ belongs to,” says Yeh.)

Still, by looking at the genes and molecular pathways that seem to be altered in perfused organs, she and her colleagues can garner some clues. At a molecular level, the team saw changes in cell pathways linked to inflammation and the structure of tissues, for example. They also saw more activity in a pathway that allows cells to remove and recycle damaged cell parts, says Yeh.

Poganik hopes to develop some kind of test that would determine which organs, on the basis of their biological age, are suitable for transplantation. He and his colleagues are also experimenting with potential drug treatments that might push the biological age of an organ even lower.

In the meantime, any liver that is not from a “perfect, young, brain-dead donor” could probably benefit from perfusion, says Yeh. The devices are already transforming transplant surgery. Just a few years ago, she says, she and her colleagues would avoid using livers from people who’d suffered a circulatory death (when the heart stops beating and there’s a damaging lack of blood flow to organs) and were over 40. Today, they use livers from such donors over the age of 70. “Perfusion has completely changed the landscape of transplantation in the last three years,” she says.

  • ✇MIT Technology Review
  • The Download: AI’s real extinction threat and age-reversal tech for eyes Thomas Macaulay
    This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Roundtables: could AI really kill us all? Employees at the world’s leading AI labs are saying there’s a real possibility that advanced AI could destroy humanity. Are they right? Or is this more scaremongering and hype? Join MIT Technology Review executive editor Niall Firth, senior AI editor Will Douglas Heaven, and AI reporter Grace Huckins for a sub
     

The Download: AI’s real extinction threat and age-reversal tech for eyes

14 September 2026 at 20:10

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.

Roundtables: could AI really kill us all?

Employees at the world’s leading AI labs are saying there’s a real possibility that advanced AI could destroy humanity. Are they right? Or is this more scaremongering and hype?

Join MIT Technology Review executive editor Niall Firth, senior AI editor Will Douglas Heaven, and AI reporter Grace Huckins for a subscriber-only conversation unpacking the debate around AI extinction. They’ll explore where the fears come from, whether they hold any water and, if they do, what we should do about them.

Register now to attend on Tuesday, September 15 at 16:00 BST / 11:00am EST / 8:00am PST.

Want to join the conversation? Subscribe to MIT Technology Review for exclusive access to all our Roundtables.

This geneticist’s age-reversal tech could help restore sight

Yuancheng (Ryan) Lu is obsessed with aging. And with eyes. As he steps outside the Whitehead Institute in Cambridge, Massachusetts, his aviator glasses darken automatically in the sun. Age-related blindness runs in his family, and his own 23andMe test came back with a mutation for macular degeneration, a top cause of vision loss in old age.

That obsession extends to his work. Lu is behind one of the coolest results in rejuvenation science and eye research: an age-reversal technique called reprogramming that repaired the optic nerves of blind mice, restoring their vision. Now, nearly the same genetic therapy he developed as a student has entered human clinical trials.

Learn more about Lu’s work on restoring sight with age-reversal therapy.

—Antonio Regalado

Yuancheng (Ryan) Lu is one of the biotechnology and health honorees on our 35 Innovators Under 35 list for 2026. Meet the rest of them here, or explore the full list across the biotechnology and health, AI, computing and robotics, and climate and energy categories.

The must-reads

I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.

1 Dario Amodei, Sam Altman, and Elon Musk have called for an AI slowdown
In a rare show of unity, the rivals agreed that AI needs stronger brakes. (Guardian) 
+ Amodei wants independent monitors and new industry-wide rules. (BBC)
+ Altman called for pacing, but not stopping. (Bloomberg $)
+ While Musk said on X that “Dario is right.”(WSJ $)
+ AI-linked stocks slumped in response. (FT $)
+ Chinese state media blasted the calls as a “Cold War” tactic. (Reuters $)
+ AI’s impacts are getting harder to predict. (MIT Technology Review)

2 Trump and Congress are resisting calls for stronger AI regulation
Trump downplayed AI risks, prioritizing the AI race with China. (NPR)
+ While the House Speaker said Congress won’t lead on AI regulation. (Politico $)
+ But Democrats are pushing for new rules before the midterms. (CNBC)
+ States and the White House are dividing over AI. (MIT Technology Review)

3 China plans to lead AI development across the BRICS countries
President Xi proposed open-source AI cooperation. (CNBC)
+ Beijing’s spy agency has warned of AI threats to national security. (FT $)

4 South Korea has tightened espionage laws to protect its chip secrets
Foreign spies can now face up to 30 years in prison. (FT $)
+ The changes follow alleged transfers of Samsung tech to China. (Reuters $)

5 The US and Mexico are teaming up to zap drones at the border
The operation may employ high-energy lasers.(Wired $)
+ Ukraine is a Wild West market for drone data. (MIT Technology Review)

6 AI agents are creating a new problem for the criminal justice system
The law has no clear answer when AI agents act independently. (Bloomberg $)
+ While courts face a flood of AI-generated lawsuits. (MIT Technology Review)

7 A Waymo pulled over and alerted police after detecting a gun
The riders were juveniles carrying a loaded AR-style ghost gun. (LA Times $)

8 Meta has been sued over data used to train its smart glasses
It allegedly used Facebook and Instagram photos without consent. (Wired $)

9 A hidden crypto farm in Mexico has put a spotlight on cartel funding
Authorities are investigating whether it stole power from a nearby dam. (Reuters $)

10 StarCraft is returning in 2030 as an open-world shooter
Fifteen years since its last release, the iconic franchise will be reborn. (Verge)

Quote of the day

“Dr. Frankenstein is telling us the monster is escaping; help us stop this.”

—Sen. Ruben Gallego, D-Ariz, calls for new AI regulation on CNN’s “State of the Union.”

One more thing


Inside the hunt for the most dangerous asteroid ever 

As asteroid 2024 YR4 hurtled toward Earth, astronomers determined that this massive rock posed a higher risk of impact than any object of its size in recorded history. Then, just as quickly as history was made, experts declared that the danger had passed. 

This is the inside story of the network of global scientists who found, followed, planned for, and finally dismissed the most dangerous asteroid ever found—all under the tightest of timelines and with the highest of stakes. Find out how they did it. 

—Robin George Andrews

We can still have nice things

A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.)

+ This master paperboy delivers newspapers with astonishing speed and skill.
+ Public Enemy and Led Zeppelin collide in this gloriously unlikely musical mashup.
+ Dozens of synchronized lasers have created extraordinary kaleidoscopic starburst patterns.
+ Check out the breathtaking winning images from the 2026 International Aerial Photographer of the Year competition.

Reading the Whole Heart: Latent-Attention Masked Autoencoders for Multimodal Cardiac Representation Learning

arXiv:2609.12035v1 Announce Type: new Abstract: Cardiovascular diagnosis rests on integrating complementary modalities, like ECG, echocardiography, chest radiographs, and clinical variables, each capturing distinct but correlated aspects of cardiac physiology. Yet most medical foundation models remain modality-specific, combining modalities only for finetuning or post-training. This discards the cross-modal evidence clinicians naturally integrate and ignores the structure within each modality. We introduce Latent-Attention Masked Autoencoders (LAMAE), a multimodal, structure-aware masked autoencoder that jointly learns patient-level representations during self-supervised pretraining. Rather than fusing modalities post hoc, LAMAE exchanges information directly in the latent space through a shared latent-attention module operating over a study-view-entity hierarchy, enabling aggregation of variable observations and graceful handling of missing modalities. Pretrained on over 1.2 million MIMIC-IV hospital stays, LAMAE outperforms modality-specific pretraining and strong contrastive and vision-language baselines across multimodal hospital-stay tasks, such as in-hospital mortality, ICD-10 and DRG coding, and length of stay, while remaining competitive on unimodal tasks. These gains persist even when only a single modality is available at test time, showing that modeling both intra- and inter-modal structure yields more robust, transferable representations.

Soft Symbol Grounding for Prototypical Concepts

arXiv:2609.12247v1 Announce Type: new Abstract: Neuro-symbolic models are usually trained with supervision only on final labels, leaving the intermediate concepts unobserved. Since many concept assignments are consistent with a given label, training can predict labels correctly while recovering the wrong concepts, a failure known as a reasoning shortcut. Prototypical networks reduce shortcuts by anchoring each concept to a few labeled examples, but existing methods still couple perception and reasoning through a hand-crafted, task-specific differentiable loss that must be redesigned for every task. We introduce \textbf{Soft-PNet}, which removes this loss: it reframes concept grounding as a Metropolis walk over a precomputed cache of feasible symbolic solutions, guided by a prototype distribution built from a single labeled anchor per concept, and trains against one KL objective between the prototype-weighted cache and the network's concept predictions. The objective is identical across tasks and remains applicable when the solution space cannot be enumerated. On \texttt{MNIST-EvenOdd}, Visual Sudoku, and \texttt{Kand-Logic} under scarce supervision, Soft-PNet matches loss-engineered prototypical networks at the concept and label levels and recovers concepts that soft-grounding baselines miss, with no loss engineering and lower training time.

AIM: A Privacy-Aware Interoperable Memory Framework for Multi-Agent Multi-User LLM Systems

arXiv:2609.12320v1 Announce Type: new Abstract: Traditional large language models (LLMs) are scoped to individual user sessions, limiting their knowledge to a single conversation and preventing them from learning user preferences that evolve over time. Existing agentic memory systems address this limitation but generally operate at the individual-user level, restricting the public knowledge that could be shared across users to improve downstream responses. We introduce AIM (Agentic Interoperable Memory), a unified, privacy-aware memory framework that enables multi-agent, multi-user LLM systems to persistently manage private and shared memory. AIM dynamically classifies information as private, scoped to one user and inaccessible to others, or public, accessible to all users. It enforces index-level access controls so that private memories are retrievable only by their owner, protecting sensitive data while allowing beneficial shared knowledge to improve coordination and consistency. We also introduce MUMBench (Multi-User Memory Benchmark), a dataset of multi-user interactions containing private and shareable information across four domains. To our knowledge, MUMBench is the first public dataset designed to evaluate multiple memory operations, including retrieval, creation, update, and deletion, in a multi-user environment. Across three independent runs on MUMBench, AIM achieves 96.0% visibility classification accuracy, 58.8% strict operation accuracy, and 70.5% state-aware operation accuracy.

BlueLM-GUI Technical Report: A Real-Device-Centric Flywheel for Self-Improving Mobile GUI Agents

arXiv:2609.12394v1 Announce Type: new Abstract: Mobile GUI agents are shifting from multi-module frameworks to native models trained end-to-end, yet industrial deployment faces three persistent gaps. Sandbox training produces a distribution mismatch with production environments; expensive real-device failures remain underutilized; and fixed benchmarks saturate, losing the power to guide iteration. We present BlueLM-GUI, a 35B-A3B mobile GUI agent built as a real-device-centric flywheel that closes these gaps through three principles. Every Sample Matters: a dual-track pipeline with Heterogeneous Triple-System Consensus evaluation and an Error Correction \& Derivation Module salvages every trajectory into usable supervision. Every Rollout Is Real: a three-stage recipe---continual pre-training, supervised fine-tuning, and agentic reinforcement learning on hundreds of real phones---grounds every rollout in real production environments, so the capability the model learns transfers directly to deployment. Every Query Evolves: a quota-driven benchmark methodology with three orthogonal axes enables precise attribution and allows the benchmark to be systematically upgraded as the model improves. BlueLM-GUI achieves 87.4 on MobileGUI-VBench, surpassing the best closed-source model by 5.1 points, and 84.9 on AndroidWorld, the best result among open-source models and competitive with closed-source models. These results demonstrate that grounding model training and iterative improvement in both real devices and the three Every principles yields strong, robust, and transferable mobile GUI capability.

Is Gaussian Splatting Becoming Neural Again? A Taxonomy and Controlled Study of Learned Parameterization

arXiv:2609.12395v1 Announce Type: new Abstract: Three-dimensional Gaussian Splatting (3DGS) combines explicit primitives with efficient rasterization, yet recent systems increasingly use neural networks to generate or share Gaussian parameters. We characterize this trend along five axes: attribute decoding, spatial sharing, view-conditioned decoding, topology generation, and amortized inference. An analysis of 19 representative methods shows that these choices address different limitations and cannot be reduced to a binary neural label. We also isolate three forms of neural parameterization in a controlled mip-NeRF 360 study. Sharing appearance and opacity improves reconstruction quality, while decoding geometric structure offers no further gain. The evidence favors selective neuralization: shared functions help when they capture reusable correlations without sacrificing the local geometric freedom of explicit splats.

When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration

arXiv:2609.12482v1 Announce Type: new Abstract: We aim to characterise the value of artificial intelligence in the workplace. Current studies largely measure this value in terms of the current automation capabilities and public adoption of AI. However, such metrics ignore the greater impacts of human--agent collaboration in transforming the nature of work. To account for this, we must expand the scope of our analysis beyond atomised tasks of today, and instead focus on how AI can augment entire workflows of the future. To ground this analysis, we establish a precise definition of AI augmentation comprising six conditions, spanning durable net value, meaningful human control, accountability and recovery, and long-term human development through learning, career pathways, and job purpose. We elaborate on these conditions and apply the framework in a case study of AI-mediated social surveys. We conclude by outlining how organisations, researchers, and government leaders can use this framework to make sense of the future of work.

Information Specialization and Constrained Synthesis in Multi-Agent LLM Forecasting: A Prospective Live-Study of the 2026 FIFA World Cup

arXiv:2609.12495v1 Announce Type: new Abstract: Large language models are being organized into multi-agent systems with specialized roles, but whether such specialization produces distinct forecasts and whether subsequent synthesis improves utility remains unclear. In this study, we carried out a live, prospective evaluation over the final 56 matches of the information-dense 2026 FIFA World Cup, keeping a frontier foundation model constant while assigning two primary forecasting agents contrasting specialist roles: a quantitative specialist focusing on structured performance statistics and a news specialist focusing on current injuries, tactics and information from press conferences. Their forecasts were then reviewed by a separate critic before being combined by a meta-agent, resulting in a sequential four-agent model. Forecasts from the betting market served as an external benchmark. The news specialist obtained the highest mean probability-weighted Top-3 utility and matched the betting market in Top-3 exact-score hits. Nevertheless, the two specialist forecasters agreed on at least two of the three scorelines in 50 out of 56 matches, and the meta-agent never generated more than one scoreline outside the specialists' forecast set. These findings show that rapidly changing, unstructured information can provide a valuable forecasting signal alongside structured statistics, whereas adding critic and meta-agent stages does not necessarily create complementary information or improve on the strongest specialist.

Beyond Generation and Accuracy: Diagnosing and Enhancing Visual Chain-of-Thought for Geometry Problem Solving

arXiv:2609.12606v1 Announce Type: new Abstract: While multimodal reasoning has advanced rapidly, solving complex geometry problems critically hinges on active visual assistance, such as constructing auxiliary lines, spurring the rise of Visual Chain-of-Thought (VCoT). However, existing evaluations typically assess visual generation quality and final answer accuracy in isolation, failing to examine whether intermediate visual aids are geometrically valid, effectively utilized in subsequent reasoning, or causally responsible for task success. To bridge this gap, we introduce GeoVAD-Bench, a diagnostic benchmark that pairs a fine-grained five-dimensional trajectory diagnosis covering perception, auxiliary quality, utilization, deductive reasoning, and final correctness with controlled No-Aux, Auto-Aux, and GT-Aux intervention settings to systematically isolate intermediate error modes, the causal gains of visual aids, and the resulting autonomy gap. Our findings reveal that while high-quality auxiliary aids offer substantial theoretical gains for geometric problem solving, autonomous generation is frequently hampered by compounding errors across geometric perception, faithful visual manipulation, visual-state grounding, and deductive reasoning. Guided by these diagnostic insights, we establish a specialized data construction pipeline encompassing geometric perception, diagram editing, and interleaved visual-textual reasoning trajectories, and develop a progressive SFT and multimodal RL training framework. The resulting model, GeoWeave-8B, outperforms the base model by +25.3% in final geometric accuracy and achieves a +30.4% gain in process average across the four intermediate diagnostic dimensions.

SteerDuplex: Steerable Duplex Speech Dialogue Models

arXiv:2609.12623v1 Announce Type: new Abstract: Full-duplex spoken dialogue models support low-latency turn taking, interruption handling, and backchanneling, yet a key capability remains underexplored: steerability, the ability to reliably shift conversational behavior along attributes such as tone, persona, speaking rate, and voice style in response to user instructions. We introduce a taxonomy of text- and audio-based steerability that identifies substantial gaps in current full-duplex models. To address this gap, we introduce SteerDuplex, a Moshi-based full-duplex speech model fine-tuned on natural conversations and synthetic dialogues targeting instruction following, vocal delivery, reasoning, and duplex interaction. We further apply two-stage reinforcement learning (RL) with hybrid rewards, combining verifiable interaction checks and judge-based semantic feedback to improve timing and response continuity. To evaluate full-duplex spoken steerability, we introduce SteerBench, a benchmark with 390 spoken prompts and 1,067 human-authored binary audio and text rubrics spanning tone, persona, style/accent, and speed/length. On SteerBench, supervised training improves audio-steering average pass rate by 44.5 percentage points over the strongest evaluated open baseline. On Audio MultiChallenge, task average pass rate improves by 7 points over its strongest evaluated open baseline. RL further raises source-clean interruption response from 72.5% to 82.5% and reduces synthetic pause barge-in from 26.5% to 9%. Steering and aggregate task scores remain comparable or higher, while reward probes reveal reward hacking through incomplete responses. Our model and benchmark support systematic research on spoken steerability, with reward analysis showing why timing gains must be evaluated alongside response completeness.

Tracing and Coordinating Cross-Layer Influence for Multimodal Model Merging

arXiv:2609.12897v1 Announce Type: new Abstract: Multimodal model merging aims to consolidate task experts into a single model that retains their complementary capabilities. Most unimodal model merging methods combine expert updates within individual layers, and multimodal approaches largely follow this design. However, an expert update changes the representations passed to subsequent layers, allowing its influence to propagate across depth and affect how visual and textual information interact. When visual and language updates are combined, later updates act on inputs already modified by earlier ones, coupling their effects. This poses two challenges: (1) how to characterize the multimodal influence of individual expert updates across depth, and (2) how to jointly combine expert updates based on their multimodal influence. To address these challenges, we propose TAC-Merge for tracing and coordinating cross-layer influence in multimodal model merging. It contains two modules, i.e., multimodal influence mapping (MIM) and coupled merge control (CMC). MIM constructs graphs of update effects and uses Ricci curvature together with expert predictions to define a shared fusion objective. CMC models interactions among coefficient adjustments and jointly optimizes regional weights to synthesize one shared model. Experiments across diverse multimodal tasks demonstrate the effectiveness of TAC-Merge in consolidating complementary expert capabilities and supporting generalization to unseen tasks.

How Good Are Frontier Models at Physics? Expert Re-Grading Reveals Broken Evaluations and Near-Saturation of Leading Benchmarks

arXiv:2609.13009v1 Announce Type: new Abstract: Low reported scores on leading physics benchmarks, including those featured in the Artificial Analysis Intelligence Index (2026), suggest that frontier language models still struggle with advanced physics, a demanding test of their scientific reasoning and quantitative problem-solving abilities. Yet this impression does not always align with domain experts' experiences using these models in their work. We revisit these reported findings by evaluating frontier models on six widely used physics benchmarks and auditing them with experts, focusing on text-only problems with verifiable final answers. For each subfield of physics, faculty and graduate researchers with relevant expertise carefully review problem statements, reference solutions, and model responses to distinguish genuine model errors from grader errors, incorrect reference solutions, and ambiguous or underspecified questions. Most audited cases initially evaluated as incorrect reflect these benchmarking issues rather than errors in the models' physics reasoning. We then ask experts to address these benchmarking issues by correcting erroneous reference solutions and repairing or excluding flawed questions. We find that GPT-5.6-Sol's measured mean@4 rises from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark, while its corrected pass@4 reaches 94.4% on the 54 retained CritPt challenges. Corrected scores are computed on the retained evaluation subsets following expert review. Scores on the audited subsets of UGPhysics, PRISM-Physics, and PHYBench also rise substantially after correction. These findings suggest that current benchmarks substantially understate frontier models' ability to solve well-posed physics problems. Near-saturation on these closed-ended tasks highlights the need for more demanding, expert-validated evaluations.

SoulAuth: An Actor-native Identity Architecture and Rust Reference Implementation for Humans and Long-lived AI Actors

arXiv:2609.11258v1 Announce Type: cross Abstract: As AI systems move from transient model invocations toward long-lived actors that persist across credentials, clients, sessions, and runtime instances, identity infrastructure must answer a basic question: where should the canonical continuity boundary be placed? This paper introduces Actor-native Identity and presents SoulAuth, an open-source Rust reference implementation for Humans and long-lived AIActors. We argue that any subject that must persist under its own identity and remain independently attributable should have an ActorIdentity that is not replaced by an Account, Credential, Client, AuthSession, IdentityBinding, or runtime instance. SoulAuth therefore treats Humans and long-lived AIActors as first-class identity subjects while keeping authentication distinct from downstream authority. Methodologically, we use a Philosophical Engineering approach that translates conceptual analysis of subjecthood into identity objects, invariants, lifecycle semantics, system responsibilities, implementation boundaries, and inspectable conformance evidence. Evaluation against the fixed SoulAuth v0.1.0 artifact shows that the implementation realizes core boundaries including Human/AIActor first-class identity status, Client/Actor separation, and Authentication/Authority separation, while gaps remain in unified Credential modeling and historical attribution anchored to ActorIdentity. We therefore report partial, not full, architecture conformance.

Reality Is the Final Verifier: On Two Key Gaps in Agentic Software Engineering

arXiv:2609.12039v1 Announce Type: cross Abstract: Software development follows an implementation-verification loop in which developers or agents iteratively revise an implementation until an evaluator, such as a test suite, accepts it. The evaluator checks the implementation against a set of requirements under a model of the deployment environment. Yet even a formal proof that the implementation satisfies the requirements under the model cannot guarantee acceptable behavior after deployment. Requirements only approximate stakeholder intent, and the model only approximates the real deployment environment. We call these together - requirement gap and model gap - the two-gap framework, which unifies the main failure modes of agentic software engineer-ing: reward hacking exploits omissions in the requirements or model, while hallucination widens the gaps by fabricating requirements or environment assumptions. Because neither gap can generally be certified closed in an open, changing world, the goal shifts from closing them to continuously narrowing them. We therefore propose an assurance-revision loop that uses deployment evidence to revise the requirements, model, or evaluator when stakeholders reject the resulting behavior. We then cast assured agentic development as a resource-allocation problem over human judgment, agent capability, and compute. The two principal bottlenecks mirror the two gaps: human judgment for the requirement gap and faithful, costly evaluation for the model gap. Reality remains the final verifier: acceptable behavior under actual deployment conditions is the ultimate test, while predeployment evaluations remain proxies for it.

Extracting Dataset Mentions in Forced Displacement and FCV Documents: A Weakly Supervised Framework with LLM-Based Label Refinement

arXiv:2609.12107v1 Announce Type: cross Abstract: Development and humanitarian organizations produce and support surveys, administrative registries, and other data resources to inform research, policy, and operations, yet systematically identifying where these datasets are referenced remains difficult. Such references are dispersed across research papers, project documents, humanitarian reports, and other unstructured text, limiting both the ability to trace data use and to identify potential gaps in data availability or dissemination. We present a weakly supervised framework for adapting dataset extraction to forced displacement and Fragile, Conflict, and Violence (FCV) documents without first constructing a large manually labeled training corpus. A lightweight model trained on general research literature generates candidate dataset mentions from unlabeled domain documents, which a frontier large language model (LLM) reviews in context, validating or rejecting candidates and correcting their extraction boundaries. The resulting annotations are supplemented with targeted synthetic and contrastive examples and used to fine-tune the lightweight model for large-scale extraction. We evaluate the resulting model on an independent gold-standard benchmark of 1,706 text passages spanning research, humanitarian, and operational documents. Across the full benchmark, the model achieves 74.1\% precision and 70.5\% recall at the mention level; among passages containing dataset references, precision reaches 89.5\%. At the passage level, the model achieves 88.2\% accuracy and 88.6\% specificity in distinguishing passages with dataset references from those without them. These results demonstrate a practical approach for constructing domain-specific supervision when labeled data are limited, and provide a technical foundation for larger-scale analysis of data use and potential gaps in the displacement data landscape.

DriftSE: Speech Enhancement with Generative Drifting

arXiv:2609.12252v1 Announce Type: cross Abstract: We propose DriftSE, a novel one-step generative framework for speech enhancement formulated as a latent distribution equilibrium problem. During training, the drifting field aligns the generator's pushforward distribution with the clean speech manifold through drifting in a latent domain. During inference, the drifting process is discarded, enabling one-step generation. We establish that its enhancement quality depends fundamentally on the choice of latent representation. Semantic latents preserve phonetic structure but fail to capture physical acoustic cues, whereas acoustic latents reconstruct the physical signal but risk linguistic hallucination. Therefore, we introduce dual-latent drifting, performing parallel drifting in both semantic and acoustic latents to simultaneously preserve phonetic intelligibility and acoustic fidelity. Additionally, we demonstrate that DriftSE enables fully unpaired training by aligning latent distributions rather than exact point-wise targets. Consequently, DriftSE facilitates cross-dataset learning in the absence of paired noisy-clean samples. Moreover, DriftSE exhibits broad architectural flexibility across different generator backbones. Extensive evaluations on additive denoising and convolutive dereverberation demonstrate robust one-step enhancement across both offline and real-time causal settings. Notably, DriftSE achieves state-of-the-art word error rates across all four evaluated datasets while strictly operating at 1 NFE. Code and audio examples are available online.

ESTS at WMT26: Routing-Informed Expert Pruning for Model Compression

arXiv:2609.12310v1 Announce Type: cross Abstract: We describe six submissions under the team name ESTS to the unconstrained WMT26 Model Compression Shared Task for English--Simplified Chinese and English--Egyptian Arabic. We submit three compression operating points per translation direction, all derived from GPT-OSS-20B. We use task-specific routing mass to rank experts and cross-lingual routing divergence to allocate retained capacity across layers, then physically remove low-importance experts. The resulting specialists are recovery-tuned on GPT-5.1-generated synthetic translation data and further compressed by applying MXFP4 quantization to the retained expert projection weights. We additionally implement a robust inference system for the instruction-conditioned WMT26 setting, including category inference, output validation, retries, segmented fallback, and source-owned JSON reconstruction. Across our six submissions, parameter counts range from 4.186B to 7.770B and packed artifact sizes from 4.55 to 6.33~GiB. Internal xCOMET-XL evaluation using GPT-5.1 pseudo-references provides an internal comparison across the submitted compression operating points.
❌