Tool-call vectors, superposed experts, and provably hidden capabilities: twelve top-conf papers from October 3–9, 2026
Twelve first-submission arXiv papers from October 3–9, 2026, eleven accepted at NeurIPS 2026 and one published at ICML 2026, covering multi-agent scheduling, tool-call interpretability, sparse attention decoding, no-thinking evaluation, mixture-of-experts routing, hidden-capability audits, offline reinforcement learning, model-quality theory, pose-free 3D inpainting, video editing, guidance schedules, and causal virtual screening.
The twelve papers in this issue were first submitted to arXiv between October 3 and October 9, 2026, the week of announcement batches that follows the September 26 – October 2 window of the previous issue. An entry qualifies here only when its own arXiv submission record carries an explicit acceptance or proceedings statement naming NeurIPS, ICML, ICLR or CVPR.
Eleven entries are confirmed NeurIPS 2026 acceptances, one of them an oral, and the twelfth is a published ICML 2026 conference paper. No ICLR 2026 or CVPR 2026 decision landed inside this window, because those cycles fall at other points in the year. Workshop records, comments that merely name a conference, review-stage comments and rejected submissions were left out, and each entry's first-submission date was checked against its own arXiv record.
At a glance
Entries are grouped by area and, inside each section, ordered by posting date. The reading signal gives one verifiable number per paper rather than a verdict.
| Paper | Area | Status | Reading signal |
|---|---|---|---|
| SquidAgent: Parallelize Wisely, Coordinate Efficiently | Multi-agent systems | Accepted at NeurIPS 2026 | Parallelising only the layers whose predicted token budget pays for coordination lifts mean throughput from 17.2 to 38.1 words per second against Claude Code, a 2.2× gain. 1 |
| How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression | Agents and interpretability | Accepted by NeurIPS 2026 Main Poster | One direction in the residual stream controls the call-or-no-call decision; removing it suppresses 93.1% of tool calls on requests that never name a verb. 2 |
| More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding | LLM inference efficiency | Accepted to NeurIPS 2026 | Decoupling key-head from value-head counts reaches 2.2× end-to-end decoding speedup at 128K context, and the distilled student keeps 96.3% of the teacher's RULER score. 3 |
| Thinking Inertia: LLMs Keep Thinking When Told Not To | LLM behaviour | Accepted by NeurIPS 2026 | DeepSeek-V4-Flash still emits visible explicit inference on 99.9% of open-ended answers with thinking switched off. 4 |
| RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing | Interpretability | Published as a conference paper at ICML 2026 | Explanations built from sparse-autoencoder features predict expert routing at mean F1 0.492 on gpt-oss-20b, against 0.417 for the strongest token-frequency baseline. 5 |
| AI Models Can Provably Hide Arbitrary Capabilities | Safety and evaluation | Accepted at NeurIPS 2026 | A classifier encrypted inside a 2.07-billion-parameter host scores 50.9% after supervised fine-tuning, against 50.6% for a locker holding no trained capability at all. 6 |
| LASER: Latent Space Adjoint Matching for Support-Constrained Entropy-Regularized Offline RL | Reinforcement learning | Accepted at NeurIPS 2026 | One fixed hyperparameter set reaches at least 95% of the best mean success rate in 34 of 40 OGBench settings, with a paired comparison at p = 1.35 × 10⁻³. 7 |
| Why Heavy-Tailed Weight Spectra Predict Model Quality | Learning theory | Accepted at NeurIPS 2026 | A free energy computed from final-layer weights alone, with no data, tracks top-5 ImageNet accuracy at Spearman ρ = −0.71 across architecture groups. 8 |
| Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement | 3D vision | Accepted to NeurIPS 2026 (poster) | Pose-free 3D scene inpainting completes in a single 0.4-second forward pass where per-scene optimisation needs 30 minutes to five hours. 9 |
| VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling | Video editing | Accepted to NeurIPS 2026 | Routing video edits through an in-context image pair scores 3.09 overall on OpenVE-Bench, against 2.41 for the best open-source baseline. 10 |
| Opening the Black Box of Classifier-Free Guidance via Information Bottleneck | Generative models | Accepted to NeurIPS 2026 | Activating guidance when a score-statistics gain crosses a threshold reaches FID 1.41 on SiT-XL/2 with the SDE sampler. 11 |
| CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening | Scientific machine learning | NeurIPS 2026 (Oral) | Sparse-interaction screening lifts LIT-PCBA EF@1% from 6.50 to 8.56 and BEDROC80.5 from 0.075 to 0.097. 12 |
Agents and tool use
1. SquidAgent: Parallelize Wisely, Coordinate Efficiently
Area tag: Multi-agent systems and agent orchestration
arXiv: 2610.08647, first submitted October 6, 2026. 1
Authors / institutions: Yexiong Lin, Yu Yao, Zhen Fang, Bo Han and Tongliang Liu at The University of Sydney; Shanshan Ye at Mohamed bin Zayed University of Artificial Intelligence, the corresponding author; with the University of Technology Sydney and Hong Kong Baptist University. 13
Status: Accepted at NeurIPS 2026. 1
Problem: Parallel multi-agent systems often finish slower than one serial agent. The authors trace that to two costs a serial agent never pays: workers re-explore context the orchestrator already holds, and independently written outputs then have to be reconciled. 13
Method: SquidAgent decides layer by layer whether parallel execution is cheaper than serial, and measures cost in predicted output tokens rather than wall-clock seconds, because the authors find language models estimate their own output length far more reliably than they estimate duration. Every token budget is estimated in one planning step, each worker forks straight from the orchestrator's session so inherited context costs nothing, and a pre-generated convention block fixes the shared interface before workers start. 13
Comparison: Nine tasks, six of them large multi-file builds, against seven baselines that include Claude Code, SeqCV, MetaGPT, AFlow and the multi-agent system AgentConductor. 13
Results: Mean deliverable throughput rises from 17.2 to 38.1 words per second against Claude Code, a 2.2× gain, and from 19.2 to 38.1 against AgentConductor, a 2.0× gain. The margin is widest on the largest tasks, 2.8× on ArcadeBox and 2.3× on SlideKit over the strongest baseline, and the summary reports a 2.6× mean wall-time speedup over Claude Code. 13
Resources: Code is public at the project repository. 13
Loading content card…
Takeaway: A team running a multi-agent harness can test the scheduling criterion without reproducing the whole system, since the decision rule is separate from the agents. The paper states its own limits: individual token estimates can miss the real output by 2–5×, the proxy covers generation-heavy work rather than tool or retrieval latency, and the parallel-or-serial choice is made per layer rather than per task. 13
2. How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression
Area tag: Agent tool use and mechanistic interpretability
arXiv: 2610.09624, first submitted October 7, 2026. 2
Authors / institutions: Xijie Gong, Jiahao Zhang, Youcheng Sun and Lijie Hu at Mohamed bin Zayed University of Artificial Intelligence; Tingxu Han at MBZUAI and Nanjing University; Wei Song at Griffith University; Ziqi Ding at the University of New South Wales; Hanqi Yan at King's College London. 14
Status: Accepted by NeurIPS 2026 Main Poster. 2
Problem: An agentic prompt stacks role instructions, tool schemas, format templates and the user's request into hundreds of tokens, so no single variable isolates the moment the model decides to call a tool instead of answering. 14
Method: The authors reduce agentic prompts to minimal contrastive pairs in which one request verb decides the outcome: swapping an execution verb such as "write" for an analysis verb such as "discuss" reliably flips the call. They build 500 such pairs across Python, Java and C++, keep 300 for analysis and 200 held out, then use activation patching and transcoder decomposition to find what carries the decision. 14
Comparison: A norm-matched random vector as the intervention control, a random suppressor and a structural control feature for the readout, plus seven models from the Qwen, Mistral and Granite families. 14
Results: The decisive state sits at layer 24 of Qwen3-8B, and patching it recovers
<tool_call> on 100% of corrupt prompts. On native multi-turn trajectories, subtracting the vector suppresses 36.7% of baseline tool calls and adding it induces calls on 70.0% of turns that had answered in text, where a norm-matched random vector moves 1.5% and 0.0% of the same turns. On requests that imply a need for tools without naming any verb, removal suppresses 93.1% of calls, and 100% at a larger gain against 16.2% for the random vector. Across seven models, full-state patching recovers calls on 93.5–100.0% of analysis prompts. 14Resources: Code is public at the project repository. 14
Loading content card…
Takeaway: Anyone debugging why an agent fires tools at the wrong moment can treat this as a handle on the decision rather than a description of it. The scope stops at the binary call-or-no-call decision: the paper does not test whether the chosen tool is the right one or whether its arguments are valid, and the formation analysis rests on behaviourally screened coding pairs under a single scaffold. 14
Language models and generation
3. More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding
Area tag: LLM inference efficiency
arXiv: 2610.04753, first submitted October 3, 2026. 3
Authors / institutions: Noam Elata, Itay Lamprecht, Mikey Shechter and Daniel Soudry at Technion – Haifa, Israel; Daniel Ohayon at Corma and Technion; Itay Hubara at a stealth startup; Lamprecht also at Crusoe AI. 15
Status: Accepted to NeurIPS 2026. 3
Problem: Sparse attention keeps only the high-probability entries of the attention matrix, which makes the multiplication of probabilities by values cheap and leaves the query-key step as the remaining cost. Group-query attention ties key heads and value heads to the same count, so shrinking the key side also shrinks the capacity that the value side provides. 15
Method: Sparse Asymmetric Group-Query Attention decouples the two head counts, cutting key heads while keeping the value heads that carry capacity, and is paired with approximate top-N attention. The paper formalises the asymmetry in a lemma and a theorem, and adds a distillation procedure that converts an already-trained group-query model instead of requiring a retraining run. 15
Comparison: A full-attention group-query baseline at the original head configuration, with and without sparse attention, measured on Llama 3.2-1B on an H200 and Qwen2.5-1.5B on a B200, against sliding-window attention with sinks, H2O and KIVI for quality. 15
Results: At 128K context and batch size 1, the combined method reaches a 2.2× end-to-end speedup on Qwen and 2.1× on Llama; at batch size 16 the same context gives 2.2× and 2.6×. Merging key heads alone is up to 1.47× faster at 128K and batch 16. The distilled student retains 96.3% of the teacher's aggregate RULER score across 8K, 16K and 32K contexts, where sliding-window attention, H2O and KIVI fall below 50% aggregate RULER at 8K. 15
Resources: Code is public at the project repository. 15
Loading content card…
Takeaway: A serving team can test the conversion route on a checkpoint it already has, and the head-count change is separable from the sparsity method it was studied with. The quality results stop at 1.5 billion parameters, the sparse kernel used is the simple approximate top-N rather than a hardware-aware one, and the distilled models keep the teacher's architecture apart from the reduced key projections. 15
4. Thinking Inertia: LLMs Keep Thinking When Told Not To
Area tag: LLM behaviour and evaluation
arXiv: 2610.11765, first submitted October 8, 2026. 4
Authors / institutions: Dianqiao Lei at Tsinghua University; Kevin Qinghong Lin at the University of Oxford, who led the project; Pan Lu at Stanford University; Philip Torr at the University of Oxford; James Zou at Stanford University. 16
Status: Accepted by NeurIPS 2026. 4
Problem: Whether a model is thinking is usually read off a proxy: a disabled thinking mode, or the absence of a long trace. Both proxies fail, because a disabled mode can still emit extended reasoning and a long trace can be filler. 16
Method: Each response is split into a pre-answer trace and a final answer, then scored three ways: an Empty-Thinking Rate for strict answer-only compliance, an instruction-aware relevance score between question and pre-answer text, and a blinded language-model judge that labels visible explicit inference. Six language models are run under six prompting interventions across Boolean, multiple-choice and open-ended questions. 16
Comparison: Native thinking-on against native thinking-off, then four stronger no-thinking interventions from ordinary instructions up to prefix forcing, with a five-annotator human study validating the three metrics. 16
Results: Under its own disabled-thinking setting, DeepSeek-V4-Flash, a 284B model, still shows visible explicit inference in 99.9% of open-ended responses. Strict answer-only prompting raises mean Empty-Thinking Rate from near zero to roughly 40% while cutting open-ended accuracy by about 15 percentage points, and lowers the Explicit Inference Rate across the corpus from 99.69% to 62.00%. On Boolean questions the same intervention drops explicit inference from 5.15% to 0.12% with accuracy flat. Scientific and technical questions combine the highest accuracy with the lowest Empty-Thinking Rate, 0.3–36.4%. Rewriting one Boolean question to offer candidate answers lifts its Empty-Thinking Rate from 10.5% to 62.5%. 16
Resources: Code and the evaluation data are public at the project repository and website. 16
Loading content card…
Takeaway: A team that switches a thinking mode off and assumes the compute is saved now has a way to check, and the three-level split separates a stripped answer from a paraphrased one. The claims concern visible output rather than hidden computation, the suite is six models and six benchmarks under deterministic decoding, and multilingual, code, tool-use and multi-turn settings remain untested. 16
Model internals and safety
5. RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing
Area tag: Mechanistic interpretability of mixture-of-experts models
arXiv: 2610.11775, first submitted October 8, 2026. 5
Authors / institutions: Ilya Lasy at the Faculty of Informatics, TU Wien; Nora Yinuo Cai, independent; Kola Ayonrinde at the UK AI Security Institute. 17
Status: Published as a conference paper at ICML 2026; the submission record reads "published as a conference paper at ICML 2026". 5
Problem: Interpretability work on mixture-of-experts models assumes each expert specialises in one coherent domain, and that assumption has not produced explanations that hold up. 17
Method: The authors propose that an expert instead covers a disjoint union of fine-grained features, and build RouterInterp on it: sparse-autoencoder features most predictive of a routing decision are selected by gradient attribution, and each expert's set of features is turned into a single natural-language explanation through an automated pipeline. 17
Comparison: A token-frequency baseline, a method scoring n-grams, and Expert Impact AutoInterp, which ranks activating windows by the expert's effect on the residual stream, evaluated layer by layer on gpt-oss-20b and OLMoE-1B-7B. 17
Results: RouterInterp reaches mean F1 of 0.492 on gpt-oss-20b and 0.602 on OLMoE-1B-7B, ahead of both baselines at every layer on both models, against 0.417 mean F1 for the strongest token-statistics baseline on gpt-oss-20b, a gain of about 18%. The paper summarises the same comparison as roughly 65% higher mean F1 than the token-frequency baseline on gpt-oss-20b. A routing-prediction probe on gpt-oss-20b reaches a macro-F1 of 0.73. 17
Resources: The paper names a repository at github.com/ilyalasy/routerinterp, which did not resolve when this issue was assembled; it also builds on public sparse autoencoders released for gpt-oss-20b. 17
Takeaway: An interpretation of a routing decision is only as good as the sparse autoencoder behind it, and the paper says so: explanation quality tracks reconstruction error across layers, and explanations grow longer as more features are added. 17
6. AI Models Can Provably Hide Arbitrary Capabilities
Area tag: Model safety and capability evaluation
arXiv: 2610.13928, first submitted October 9, 2026. 6
Authors / institutions: Andis Draguns at Contramont Research; Sumeet Ramesh Motwani at the University of Oxford; Raymond Douglas at ACS Research; Christian Schroeder de Witt at the University of Oxford. 18
Status: Accepted at NeurIPS 2026. 6
Problem: Capability evaluations assume a hidden ability can be brought out by enough fine-tuning, noise or trigger search. If that assumption fails, passing an audit says nothing about what a model can do. 18
Method: The authors embed an encrypted neural network inside a host model's weights using digital lockers, so the hidden circuit decrypts and runs only on a secret trigger. Under the stated cryptographic assumptions the circuit stays hard to elicit or interpret even with full access to the weights, and the construction is implemented in PyTorch. 18
Comparison: A password-locked model trained on the same tasks, and a control whose locker holds a random network with no trained capability, tested with supervised fine-tuning, latent adversarial training, weight noising at two scales, and gradient-based trigger search. 18
Results: On chemical reaction prediction, a host with 38 layers, about 2.07 billion parameters and a maximum residual-stream width of 10,266 reaches 73.5% accuracy when triggered. Under supervised fine-tuning it scores 50.9%, against 50.6% for the random-network control, and under latent adversarial training 52.9% against 53.6%. The password-locked baseline of the same task reaches 92.7% when triggered and is fine-tuned to 100% accuracy on the subset-parity task, where the compiled model stays at chance. Gradient-based search reaches 52.7% on a 16-bit trigger. 18
Resources: The authors release the models used in the experiments; the record lists no code repository. 6
Takeaway: Elicitation-based auditing cannot be the guarantee for a model bought from someone else, and the paper is careful about what it has shown. The circuits are deliberately constructed rather than claimed to arise from ordinary training, the single-position execution unit carries a quadratic width penalty that would stop a large circuit fitting a fixed-size host, and full integration into a pretrained model remains future work. 18
Reinforcement learning
7. LASER: Latent Space Adjoint Matching for Support-Constrained Entropy-Regularized Offline RL
Area tag: Offline reinforcement learning
arXiv: 2610.08989, first submitted October 6, 2026. 7
Authors / institutions: Songyuan Zhang, Oswin So, Eric Yang Yu and Chuchu Fan at MIT; Matthew Cleaveland and Peter Crowley-Dolen at MIT Lincoln Laboratory. 19
Status: Accepted at NeurIPS 2026. 7
Problem: Offline reinforcement learning is limited by actions the dataset never contains. Learning a flow-matching behaviour policy and optimising inside its latent space contains those actions, but the latent policy then tends to collapse onto one mode or to exploit sharp edges of the learned critic. 19
Method: LASER adds entropy regularization in the latent space and optimises it by adjoint matching, which reaches the optimal latent policy without backpropagating through the whole generation trajectory. Support constraints and entropy regularization are treated as separate jobs: the first keeps the policy on the data's actions, the second keeps it from collapsing inside that support. 19
Comparison: Six offline reinforcement learning baselines with flow or diffusion policies, including ReFORM, QAM, IFQL and DSRL, across 40 OGBench settings on clean and noisy datasets. 19
Results: With one fixed hyperparameter set for every task, LASER reaches at least 95% of the best mean success rate in 34 of the 40 settings and a mean success rate of 100% in 13 of them, with a paired statistical comparison at a global p of 1.35 × 10⁻³. Adjoint matching trains 3.6× faster than backpropagation through time, 123 against 34 iterations per second, and the ablation over the decoder prior leaves final success rates comparable, which the authors read as the entropy and behaviour regularizers acting separately. 19
Resources: Code and demonstrations are released through the project website. 19
Takeaway: A practitioner choosing an offline algorithm can take the fixed-hyperparameter result as the headline, since the baselines keep their per-task tuning. The costs the paper names are real: enforcing the support constraint through the decoder means inheriting the decoder's own out-of-distribution errors, and sampling from the flow model during training stays expensive for deep neural ordinary differential equations. 19
Theory and optimization
8. Why Heavy-Tailed Weight Spectra Predict Model Quality
Area tag: Learning theory and model-quality prediction
arXiv: 2610.13123, first submitted October 9, 2026. 8
Authors / institutions: Joseph Wilson and Fred Roosta at the University of Queensland; Chris van der Heide and Liam Hodgkinson at the University of Melbourne; Zhichao Wang at ICSI and the University of California, Berkeley; Michael W. Mahoney at ICSI, Lawrence Berkeley National Laboratory and the University of California, Berkeley. 20
Status: Accepted at NeurIPS 2026. 8
Problem: Heavy-tailed spectra in trained weight matrices predict how good a model will be, and nobody has explained why. Data-dependent quality metrics cannot fill the gap either, because they are expensive on large datasets and unavailable when the training set is not released. 20
Method: The authors derive the Bayes free energy in a form computable from the final layer's weights alone, a data-free quantity that measures the marginal likelihood. Assuming the weight spectra follow the heavy-tailed High-Temperature Marchenko–Pastur law, they take the limit of that free energy, show it is convex in a heaviness parameter, and locate the optimal heaviness. 20
Comparison: The established heavy-tailed self-regularization metrics on the same checkpoints, across torchvision convolutional networks including ResNet, VGG and DenseNet, Pythia models from 70M to 12B parameters, and the Qwen3.5 suite. 20
Results: Across the torchvision groups, the metric tracks top-5 ImageNet accuracy at a Spearman correlation of −0.71 on the pooled fit, with per-architecture correlations of −1.00 for ResNet, −0.92 for RegNet and −0.80 for ConvNeXt and DenseNet; the sign is negative because a lower free energy means a better model. On Pythia the correlation with LAMBADA accuracy is −0.66, and on Qwen3.5 with MMLU-Pro it is −0.77. 20
Resources: The record lists no code repository; the models studied are public checkpoints. 8
Takeaway: A team that cannot see the training data of a model it is handed can now rank models by a quantity it can compute from the weights, and the derived optimal heaviness gives a reference point during training. The derivation covers final-layer weights and regression, the dual form cannot be compared across models with different activation dimensions, and neither Vision Transformers nor dropout models showed a correlation. 20
Vision and 3D
9. Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement
Area tag: 3D scene reconstruction and inpainting
arXiv: 2610.11857, first submitted October 8, 2026. 9
Authors / institutions: Jingyi Pan and Qiong Luo at The Hong Kong University of Science and Technology (Guangzhou) and HKUST; Dan Xu at HKUST, the corresponding author. 21
Status: Accepted to NeurIPS 2026 (poster). 9
Problem: Filling in a removed object inside a 3D scene usually assumes accurately calibrated cameras, which rules out the casual, unposed photo collections where this kind of editing is wanted and adds a registration step before it can start. 21
Method: FreeInpaint extends a feed-forward 3D foundation model to masked inputs, so pose estimation, geometry recovery and inpainting happen together. A learnable mask attention lets masked regions pick up context only in deeper layers while reliable observations keep their spatial anchoring, and support token refinement adds diffusion-generated evidence as confidence-weighted auxiliary tokens where one forward pass lacks the appearance information to complete a region. 21
Comparison: Per-scene optimisation methods on SPIn-NeRF, 360-USID and LLFF, plus two cascaded feed-forward pipelines built by pairing the 2D inpainting methods LAMA and MVInpainter with a pose estimator. 21
Results: Inference takes 0.4 seconds in a single forward pass, where the optimisation-based methods need 30 minutes to five hours. On 360-USID, replacing a plain or hard mask with the learnable mask attention raises PSNR from 16.96 and 17.97 to 18.58 and lowers absolute trajectory error from 0.0462 and 0.0377 to 0.0340, working from 4 unposed images on SPIn-NeRF and LLFF and 8 on 360-USID. 21
Resources: A project page carries demonstrations; the record lists no code repository. 9
Takeaway: The speed divides the two halves of the method usefully: pose-free reconstruction is the feed-forward part, and the refinement loop is where you spend more if the occluded region still looks thin. What bounds the result is the 2D diffusion prior doing the fill, and the paper reports error accumulating in 360-degree scenes seen from only one or two views. 21
Video and generative models
10. VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling
Area tag: Video editing
arXiv: 2610.12104, first submitted October 8, 2026. 10
Authors / institutions: Leigang Qu, Wei Chow, Yicong Li, Wenjie Wang and Tat-Seng Chua at the National University of Singapore; Feng Cheng, Ziyan Yang, Bangbang Yang, Zhaoyang Huang and Yan Zeng at ByteDance Seed; with the University of Science and Technology of China. 22
Status: Accepted to NeurIPS 2026. 10
Problem: A video editor needs source, instruction and edited video triplets, which are expensive to annotate and hard to synthesise at scale, while image editing already has millions of such pairs. 22
Method: Video editing is decomposed into a chain of composable sub-tasks that routes the editing intent through the image domain, so image and video corpora can be trained together under one diffusion objective. An image editing pair, supplied by the user or produced by the model, is prepended as an in-context demonstration and serves as the spatial appearance blueprint for every output frame. A dual-frequency 3D rotary position encoding puts the image demonstration and the video frames into one coordinate system, and Chain-of-Editing runs the sub-task chain as progressive diffusion stages so more test-time compute buys more fidelity without retraining. 22
Comparison: OpenVE-Bench across eight editing categories, scored by Gemini 2.5 Pro and by Seed1.6-VL as an independent judge, against Runway Aleph, VACE, OmniVideo, InsViE, Lucy-Edit, ICVE, DITTO and OpenVE-Edit. 22
Results: Under the independent judge, VINCIE-NExT scores 3.09 overall against 2.41 for OpenVE-Edit and 2.25 for ICVE, narrowing the gap to the closed-source Runway Aleph at 3.50. Global style at 4.25 and creative edit at 3.39 are its strengths and camera edit at 2.01 is its weakest category, matching the ordering from the other judge. The two-stage chain runs in 53.5 seconds on one H200. 22
Resources: A project page carries results, and the paper reports no released code. 10
Takeaway: The image-editing step is the whole design and also its ceiling: a hallucinated detail in the edited keyframe propagates to every frame, because stage two treats that edit as a hard spatial anchor. The paper also notes that one keyframe does not adequately describe videos with large temporal variation or big camera motion. 22
11. Opening the Black Box of Classifier-Free Guidance via Information Bottleneck
Area tag: Diffusion model guidance
arXiv: 2610.13369, first submitted October 9, 2026. 11
Authors / institutions: Jiayang Gao, Luyao Fan and Jia Wang at Shanghai Jiao Tong University; Tianyi Zheng, Fengxiang Yang and Bo Li at vivo BlueImage Lab; Jiayang Zou at Stanford University; Lv Tang at the University of British Columbia. 23
Status: Accepted to NeurIPS 2026. 11
Problem: Classifier-free guidance improves conditional generation, and when to switch it on is still chosen by hand. Fixed schedules and hand-designed dynamic ones both lack an account of when the conditioning input actually carries information. 23
Method: The paper treats the diffusion trajectory as a time-dependent information bottleneck and shows that conditional information is not equally useful throughout sampling: it becomes much more valuable after a critical stage defined by a Fisher-information balance. From that, it builds a training-free rule that tracks conditional and unconditional score statistics, estimates the guidance gain and activates guidance once that estimate crosses a threshold. Interval-based guidance falls out as a limiting case of the rule. 23
Comparison: Class-conditional ImageNet generation on SiT-XL/2 under flow matching and DiT-XL/2 under DDPM, and text-to-image synthesis on Stable Diffusion 3 and FLUX.1, scored by FID, sFID, IS, CLIP score, HPSv2 and an aesthetic score. 23
Results: On SiT-XL/2 the method reaches its best FID of 1.41 with the SDE sampler and 1.47 with the ODE sampler, and it improves on the baseline for DiT-XL/2 at both resolutions. On FLUX.1 at 256×256 it lowers FID and raises HPSv2.1 to a mean of 30.73 at a batch of 16 against 20.13 for standard guidance, with a slight reduction in CLIP score, and it lowers FID in every batch setting tested. 23
Resources: The method is a training-free wrapper around an existing sampler; the record lists no code repository. 11
Takeaway: The threshold rule is worth trying before hand-tuning a guidance interval, since it costs a few score evaluations and recovers interval guidance as a special case. The analysis covers a class of stochastic flow models, and the paper says fully characterising deterministic couplings is what remains open. 23
Scientific machine learning
12. CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening
Area tag: Scientific machine learning for drug discovery
arXiv: 2610.07340, first submitted October 5, 2026. 12
Authors / institutions: Loka Li, Jin Tian and Kun Zhang at Mohamed bin Zayed University of Artificial Intelligence; Zhang also at Carnegie Mellon University. 24
Status: NeurIPS 2026 (Oral); the submission record reads "NeurIPS 2026 (Oral)". 12
Problem: Virtual screening aligns a protein and a molecule in one shared embedding space, which mixes the few local interactions that actually govern binding with correlations that only hold for the training targets. Training data contain observed binders only, so the nuisance correlations are the easy thing to learn. 24
Method: The authors formalise sparse binding as a V-structure causal model under Heckman-style selection and prove three results: the latent concepts are not identifiable without sparsity constraints, they are component-wise identifiable under structural sparsity, and a low-rank relaxation gives subspace identifiability. CausalBind implements the principle with a concept extractor, a sparsity mask and a training loss, in three variants. 24
Comparison: Classical docking systems, structure-aware scoring models and recent retrieval-based representation learners, including LigUnity and HypSeek, on DUD-E with 102 targets and LIT-PCBA with 15 targets. 24
Results: Every variant beats the reproduced retrieval baselines, with the largest relative gains on LIT-PCBA early enrichment. There EF@1% rises from 6.50 for LigUnity and 4.62 for HypSeek to 8.56, and BEDROC80.5 from 0.075 and 0.061 to 0.097. Against the concurrent two-stage S2Drug pipeline, which keeps the best DUD-E BEDROC80.5, all three variants lead on DUD-E AUROC and EF@1% and on all three LIT-PCBA metrics. 24
Resources: Code and reproduction scripts are at the project repository. 24
Loading content card…
Takeaway: The sparsity is the contribution, and it is the thing to test on your own target set, because the gain is largest where property shortcuts are least useful. The evidence sits inside retrieval-based screening under cross-dataset generalisation, the paper does not claim uniform gains across sequence-family or time-stratified splits, and its interpretability analysis is model-level attribution rather than proof of a physical mechanism. 24
What to open first
- Running multi-agent or tool-using systems: SquidAgent (2610.08647) for a scheduling rule that needs no reinforcement learning, and the tool-call vector (2610.09624) if tools fire at the wrong moment and you want to see the decision rather than guess at it. 1
- Serving or fine-tuning language models: SAGA (2610.04753) if long-context decoding is your cost and you have a checkpoint to convert; RouterInterp (2610.11775) if you need to know what an expert in your mixture-of-experts model does. 3
- Evaluating a model you did not train: the hidden-capabilities paper (2610.13928) for why an elicitation audit cannot close the question, and Thinking Inertia (2610.11765) if you ship a thinking mode and want to know what its off switch really saves. 6
- Working on offline control: LASER (2610.08989) for the fixed-hyperparameter result on noisy datasets and the adjoint-matching training speed. 7
- Studying training dynamics: the weight-spectra paper (2610.13123) if you need a quality metric that works without the training data, and the guidance paper (2610.13369) if you want the schedule you have been hand-tuning replaced by a threshold. 8
- Building 3D or video editing tools: FreeInpaint (2610.11857) if your captures have no camera poses, and VINCIE-NExT (2610.12104) if you have image-editing data and no video-editing data. 9
- Screening molecules: CausalBind (2610.07340) if your property-debiased benchmark is LIT-PCBA-like and your baseline leans on dense alignment. 12
References
- 1SquidAgent arXiv record
arxiv.org
- 2Tool-call vector arXiv record
arxiv.org
- 3SAGA arXiv record
arxiv.org
- 4Thinking Inertia arXiv record
arxiv.org
- 5RouterInterp arXiv record
arxiv.org
- 6Hidden capabilities arXiv record
arxiv.org
- 7LASER arXiv record
arxiv.org
- 8Heavy-tailed spectra arXiv record
arxiv.org
- 9FreeInpaint arXiv record
arxiv.org
- 10VINCIE-NExT arXiv record
arxiv.org
- 11CFG arXiv record
arxiv.org
- 12CausalBind arXiv record
arxiv.org
- 13SquidAgent paper
arxiv.org
- 14Tool-call vector paper
arxiv.org
- 15SAGA paper
arxiv.org
- 16Thinking Inertia paper
arxiv.org
- 17RouterInterp paper
arxiv.org
- 18Hidden capabilities paper
arxiv.org
- 19LASER paper
arxiv.org
- 20Heavy-tailed spectra paper
arxiv.org
- 21FreeInpaint paper
arxiv.org
- 22VINCIE-NExT paper
arxiv.org
- 23CFG paper
arxiv.org
- 24CausalBind paper
arxiv.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.