1110 tokens a second, 79% of unlearned text rebuilt, and a 14B agent 0.2 points off Opus: twelve top-conf papers from October 17–23, 2026
Twelve first-submission arXiv papers from October 17–23, 2026 — nine accepted at NeurIPS 2026, two at ICML 2026 and one at CVPR 2026 — covering low-bit inference, SVD compression, reasoning-token control, test-time selection, emergence claims, agent reinforcement learning, MCP fault tolerance, forecasting, latent dynamics, unlearning privacy, world models and longitudinal radiology reports.
Twelve papers in this issue were first submitted to arXiv between October 17 and October 23, 2026, the week that follows the October 10–16 window of the previous issue. An entry qualifies only when its own arXiv submission record states an accepted NeurIPS, ICML, ICLR or CVPR 2026 venue.
Nine entries are NeurIPS 2026 papers, two of them spotlights. Two are ICML 2026 papers, one of them with a PMLR volume line in its record, and one was accepted by CVPR 2026.
No ICLR 2026 acceptance fell inside this window. The five ICLR mentions among the 207 in-window records were all ICLR 2027 submissions under review, so they were excluded along with workshop tracks, bare conference names and rejected submissions.
Of the 100 papers in this window whose records name an accepted venue, these twelve were picked to spread across directions and to keep papers that report their headline numbers. Each first-submission date was checked against the paper's 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 |
|---|---|---|---|
| BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models | Low-bit inference | Accepted at NeurIPS 2026 | 1110.65 tokens/s at a 47.91 mean score on Bonsai-1.7B, against 566.23 tokens/s at 47.65 for Bonsai plus EAGLE-3. 1 |
| Advancing SVD-based LLM Compression via Layer-Wise Error Model Search | LLM compression | Accepted at ICML 2026 | Up to 4.8 points of zero-shot accuracy over existing rank-allocation search, and 15% average perplexity from KFAC-SVD. 2 |
| SteadyThought: Mitigating LLM Under-Thinking via Thought-Level Preference Optimization | Reasoning post-training | Accepted at NeurIPS 2026 | 30.0% fewer reasoning tokens with 1.85 points more accuracy on the 1.5B model; 34.5% fewer tokens on the 8B. 3 |
| Correct but Unselectable: The Hidden Interface Tax in Multi-Candidate Reasoning | Test-time selection | Accepted at NeurIPS 2026 | GPQA Diamond final accuracy rises from 5.6% to 19.2% from the interface layer alone. 4 |
| Claims of emergence in AI should be grounded in information decomposition | Evaluation methodology | NeurIPS 2026 spotlight | 73% of the collaborative residual survives the paper's substitution test. 5 |
| BIRD-RL: Scaling Agentic Reinforcement Learning over Stateful Data-Centric Environments | Agent training | Accepted to NeurIPS 2026 main track | A trained 14B model reaches 48.00% success where Claude-Opus-4.6 as an agent reaches 48.20%. 6 |
| Tool-Calling LLM Agents Under Realistic MCP Faults: Scale Does Not Purchase Fault Tolerance | Agent robustness | Accepted at NeurIPS 2026 (poster) | Task score falls from 8.08 to 3.23 out of 10 under permanent-heavy faults, and the 120B backbone absorbs them no better than the 20B. 7 |
| Paloma: Phase-Conditioned Residual Modulation for Time Series Forecasting | Time-series forecasting | Accepted by NeurIPS 2026 (spotlight) | Best MSE on 7 of 12 datasets and best MAE on 11, averaged over four horizons. 8 |
| Meta-iLaD: Identifiable Latent Dynamics via Meta-Learning of Dynamics Environments | Latent dynamics | ICML 2026 proceedings, PMLR 306 | Out-of-distribution prediction MSE 78.7 against 108 for MONODE on held-out MoCAP subjects. 9 |
| TRACE: Data-Free Text Reconstruction Attacks against Approximate Unlearning in LLMs | Unlearning privacy | Accepted to NeurIPS 2026 | 78.75% of unlearned samples recovered at ROUGE-L recall 1.0, against 10.00% for the perplexity-only reconstruction. 10 |
| AEON: Unifying Video and 3D World Models | World models | Accepted to NeurIPS 2026 | DaVis reconstruction PSNR 24.661 against 23.929 for NeoVerse, with no pose or depth input at inference. 11 |
| TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation | Clinical imaging | Accepted by CVPR 2026 | Clinical F1 0.511 against 0.474 for the strongest prior longitudinal method. 12 |
Compression and inference
1. BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models
Area tag: Low-bit inference and speculative decoding
arXiv: 2610.28500, first submitted October 20, 2026. 1
Authors / institutions: Ning Zhang, Shihao Wang, Jinrui Zhang, Chaodong Xiao and Lei Zhang, at the Department of Computing, The Hong Kong Polytechnic University. 13
Status: Accepted at NeurIPS 2026, as stated in the arXiv record. 1
Problem: Ultra-low-bit backbones cut weight storage, memory traffic and linear-layer cost. Multi-token prediction cuts the number of sequential decode steps. Combining the two does not speed decoding at comparable quality: masked drafting lacks the intermediate tokens that next-token training supplies, and wider computation plus causal key–value updates eat the savings. 13
Method: One model is trained with a joint next-token and multi-token objective, so the same weights do autoregressive generation and block-causal drafting. Three decoding policies are offered: confidence-only, speculative-only and hybrid. Accepted-prefix alignment keeps the committed prefix, the materialized causal KV boundary and the CUDA-graph replay state in step, and the runtime reuses captured graphs over shared KV storage. 13
Comparison: SFT autoregressive baselines on BitNet b1.58-2B and Bonsai-1.7B and 4B, compact BF16 Qwen2.5-1.5B, Llama-3.2-3B and Qwen3-1.7B, an AWQ-quantized model, and a Bonsai model paired with the EAGLE-3 learned drafter. 13
Results: On an H20 at block size 8, five-task mean speedups are 2.2× over BitNet, 2.4× over Bonsai-1.7B and 2.1× over Bonsai-4B. Bonsai quality stays close to SFT, at 47.91 against 47.36 and 63.43 against 63.55. Bonsai plus EAGLE-3 reports 566.23 tokens/s at 47.65, where BitMTP hybrid decoding reaches 1110.65 tokens/s at 47.91. On an RTX 3090 the method reaches up to 2.0× mean throughput, and 3.3× on MBPP+ on a Jetson Orin. 13
Resources: Code is public at the project repository. 13
Loading content card…
Takeaway: The speedup comes from one low-bit backbone proposing and verifying, which avoids paying for a separate BF16 drafter. BitNet is the backbone that pays in quality: at block size 8 its hybrid decoding reaches 2.2× with a mean score of 43.25 against SFT's 45.95, while block size 4 with speculative-only decoding keeps 45.17 at 1.7×. 13
2. Advancing SVD-based LLM Compression via Layer-Wise Error Model Search
Area tag: LLM compression and rank allocation
arXiv: 2610.32878, first submitted October 23, 2026. 2
Authors / institutions: Moritz Thoma, Maximilian Groezinger, Maximilian Forstenhäusler, Emad Aghajanzadeh, Manoj Rohit Vemparala, Christos Anagnostopoulos, Pierpaolo Mori, Nael Fasfous, Alexander Frickenstein, Daniel Mueller-Gritschneder and Ulf Schlichtmann, across the Technical University of Munich, BMW Group and the University of Glasgow. 14
Status: Accepted at ICML 2026. 2
Problem: Low-rank SVD compression replaces dense weight matrices with smaller factors that stay fast on standard inference engines. Two pieces of it remain weak: rank is allocated globally from error proxies that do not model how error propagates between layers, and Fisher-based decompositions collapse in rank. 14
Method: LEMS builds a layer-wise error surrogate that combines local and global layer importance with a propagation bias, then allocates rank across the whole model by integer linear programming. KFAC-SVD replaces the per-sequence gradient estimator with token-wise Kronecker-factored statistics, which avoids the rank deficiency of the Fisher-based route. 14
Comparison: SVD baselines and rank-allocation search strategies, including ASVD, SVD-LLM and SVD-LLMv2, MRCS and Fisher-based estimators, across the Mistral, Qwen3 and Llama 3 families up to 70B parameters. 14
Results: LEMS improves zero-shot accuracy by up to 4.8 percentage points over existing search strategies, and the gain carries to 70B models. KFAC-SVD improves perplexity by 15% on average. 14
Resources: Code and the project page are at lems-svd.github.io. 14
Takeaway: Rank allocation and decomposition quality are separate decisions, and the paper moves both: one chooses how much rank each layer gets, the other chooses which statistics define the truncation. Both replace a step inside an existing SVD pipeline, and the public code makes the accuracy gain testable at a reader's own compression ratio. 14
Reasoning and test-time selection
3. SteadyThought: Mitigating LLM Under-Thinking via Thought-Level Preference Optimization
Area tag: Reasoning post-training and token efficiency
arXiv: 2610.25045, first submitted October 18, 2026. 3
Authors / institutions: Zhenyue Peng, Lin Yang, Jie Wang, Xiaoqi Ni, Hanzhu Chen, Xize Liang, Yanjun Wu, Rujing Wang, Kaiwen Zhou and Jianye Hao, across the Hefei Institutes of Physical Science at the Chinese Academy of Sciences, the University of Science and Technology of China, Shanghai Artificial Intelligence Laboratory, Huawei and Tianjin University. 15
Status: Accepted at NeurIPS 2026. 3
Problem: Large reasoning models jump between reasoning trajectories, and they often jump away from a promising one and never return. Suppressing the switch outright is the obvious fix, and it throws away the trajectories that would have worked. 15
Method: The paper treats under-thinking as a preference problem at the switch point. From an identified point, a single-path continuation is forced to completion; the new response becomes the preferred output and the original the dispreferred one, giving a thought-level preference objective. Training uses mathematical data only. 15
Comparison: Vanilla models and four test-time efficiency methods — NoThink, NOWAIT, SEAL and TIP — on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B, 14B and 32B, scored on MATH-500, AIME 2024, GSM8K and LiveCode. 15
Results: Average token consumption falls by 30.0%, 34.5%, 19.3% and 20.4% on the 1.5B, 8B, 14B and 32B models, while average accuracy rises by 1.85, 4.00, 2.30 and 2.93 points. On the 1.5B model the overall score moves from 55.43 to 57.28 while tokens drop from 6,682 to 4,680, where NoThink scores 37.20. On LiveCode, a coding benchmark absent from training, tokens fall by 23.8%, 20.8%, 21.0% and 15.9% and accuracy rises by 2.9, 11.4, 4.7 and 3.3 points. The largest single-dataset reduction reported is 63.8%. 15
Resources: Code is public at the project repository. 15
Loading content card…
Takeaway: Steady Thought cuts tokens and raises accuracy at the same time, where NoThink, the bluntest suppression baseline, loses 18.23 points overall while cutting 81.3% of the tokens. The gains carry to a benchmark outside the training domain, which is the evidence that what transfers is the switching behaviour itself. 15
4. Correct but Unselectable: The Hidden Interface Tax in Multi-Candidate Reasoning
Area tag: Test-time scaling and answer selection
arXiv: 2610.32235, first submitted October 23, 2026. 4
Authors / institutions: Qianli Ma, Zhiqing Tang, Fanshuai Meng, Zhi Yao, Hanshuai Cui and Weijia Jia, at Beijing Normal University and the Beijing Normal-Hong Kong Baptist University. 16
Status: Accepted at NeurIPS 2026. 4
Problem: A small model deployed under cost, latency or privacy limits cannot fall back on a larger one when reasoning fails, so generate-then-rerank is its natural test-time scaling route. More candidates do not reliably raise accuracy: the right answer can sit in the pool and still be unusable, because it is incomplete, unparseable, or written in a different answer mode than the fixed selector expects. 16
Method: The paper names that loss the hidden interface tax and splits it into oracle coverage, meaning a correct candidate exists, and selectable coverage, meaning the selector can consume it. Its fix, Compatibility-Aware Candidate Construction, is an interface layer between the candidate source and the fixed selector: it repairs fragments, aligns answer modes and removes selector-confusing noise, retraining neither side. 16
Comparison: Majority voting and three LLM judging variants on the same candidate pools, against self-consistency-style candidate expansion and fine-tuning baselines, on GSM8K, competition_math, GPQA Diamond and MMLU-Pro. 16
Results: On GPQA Diamond the final accuracy goes from 5.6% to 19.2%, 13.6 points, from the interface layer alone; with proposer-side strengthening it reaches 25.8%, 20.2 points up. On GSM8K the clean oracle coverage moves from 0.4882 to 0.6391. On MMLU-Pro the same construction raises oracle coverage while final accuracy under the original selector falls 0.85 points, and only the proposer-side change brings it back up. 16
Resources: Code is public at the project repository. 16
Loading content card…
Takeaway: The MMLU-Pro result is the useful part for anyone running this pipeline: a candidate pool made more answer-like can feed a fixed selector worse, so coverage and conversion have to be measured separately before adding samples. The GPQA gain shows how much accuracy a weak selector can strand. 16
What evaluations measure
5. Claims of emergence in AI should be grounded in information decomposition
Area tag: Evaluation methodology and information theory
arXiv: 2610.28479, first submitted October 20, 2026. 5
Authors / institutions: Christoph Riedl at the Khoury College of Computer Sciences, Northeastern University, and Fernando E. Rosas at the Department of Informatics, University of Sussex. 17
Status: NeurIPS 2026 spotlight; the record's journal reference names the 40th Conference on Neural Information Processing Systems. 5
Problem: A language model shows a capability its components lack, a multi-agent system beats each agent, a human–AI pair solves what neither solves alone. Researchers call each of these emergence or synergy, and the labels are informal enough that two claims cannot be compared or disputed. 17
Method: A position paper arguing that partial information decomposition supplies the missing formalism, because it separates unique information held by one component, redundant information shared across components, and synergistic information available only from the parts together. The proposed criterion: behaviour is emergent only when the system carries synergistic information about task-relevant outcomes. 17
Comparison: The framework is applied at three scales — attention heads in a language model, autonomous agents, and human–AI teams — against the informal emergence claims each literature currently makes. 17
Results: In the paper's reanalysis of human–AI team data, users switch from an incorrect solo answer to a correct joint one on 17–19% of the easiest items and 31–38% of the hardest. A substitution test leaves 73% of the collaborative residual standing. Collaborative residuals are under-dispersed (φ = 0.658), and user-by-difficulty interactions add no predictive value (ΔELPD = −1.7) at 80% power to detect non-linearity as small as 16% of the additive effect. 17
Resources: Code is public at the project repository. 17
Loading content card…
Takeaway: The proposal makes an emergence claim checkable: name the information atoms a system carries about the outcome, and the claim becomes a measurement with a falsifiable prediction. The numbers above come from reanalysing behavioural data that already existed, so they show what the framework extracts from a study designed without it. 17
Agent training and agent robustness
6. BIRD-RL: Scaling Agentic Reinforcement Learning over Stateful Data-Centric Environments
Area tag: Agentic reinforcement learning
arXiv: 2610.29426, first submitted October 21, 2026. 6
Authors / institutions: Ge Qu, Jinyang Li, Xiaohan Xu, Xiaolong Li, Bowen Qin, Nan Huo, Ziwei Tang, Chenhao Ma and Reynold Cheng, at The University of Hong Kong, the National University of Singapore and The Chinese University of Hong Kong, Shenzhen. 18
Status: Accepted to the NeurIPS 2026 main track. 6
Problem: Agentic RL frameworks are built around tools, and a stateful task such as SQL debugging breaks that shape. Proper isolation wants a separate database container per rollout, which is expensive; reusing a database without tracking its evolving state contaminates trajectories and destabilises policy updates. 18
Method: A four-stage curriculum that alternates foundational agent reasoning with trajectory-adaptive on-policy exploration, avoiding a direct end-to-end jump. Underneath it, a trajectory-scoped persistent execution infrastructure reuses database replicas while keeping each trajectory's state isolated. Models are Qwen2.5-Coder-7B and 14B. 18
Comparison: BIRD-CRITIC-SQLite, against general-purpose models including GPT-5.4-Pro, Claude-Opus-4.6, Claude-Sonnet-4.5, Gemini-3.1-Pro, GLM-4.7, Kimi-K2.5 and Qwen3-Coder-480B, and against multi-turn agents including Claude-Opus-4.6 and the supervised BIRD-FIXER-7B and 14B. 18
Results: The trained 7B model reaches 44.40% overall success and the 14B model 48.00%, against 48.20% for the Claude-Opus-4.6 agent, 48.80% for Gemini-3.1-Pro, 42.00% for GPT-5.4-Pro and 40.00% for BIRD-FIXER-14B. The infrastructure change raises spatial RL efficiency 12.8×, and the setup needs no Kubernetes. The same models also handle SQL generation. 18
Resources: Code is public at the project repository. 18
Loading content card…
Takeaway: A 14B model trained this way lands 0.2 points behind the strongest frontier agent in the table on this benchmark. The enabling change is infrastructure for trajectory isolation plus a staged curriculum, which is the part that transfers to other stateful environments. 18
7. Tool-Calling LLM Agents Under Realistic MCP Faults: Scale Does Not Purchase Fault Tolerance
Area tag: Agent robustness and evaluation
arXiv: 2610.32555, first submitted October 23, 2026. 7
Authors / institutions: Antonio Ken Iannillo at SnT, University of Luxembourg, Joshua Segun Owotogbe and Indika Kumara at Tilburg University, Roberto Natella at the Gran Sasso Science Institute, Francesco Avallone at the University of Naples Federico II, and Kristina Kudryavtseva at SnT, University of Luxembourg. 19
Status: Accepted at NeurIPS 2026 as a poster. 7
Problem: Agents increasingly reach their tools through the Model Context Protocol, and the benchmarks that measure them assume tools answer correctly and on time. What happens to an agent when a tool fails permanently, briefly or intermittently is untested. 19
Method: MCP-Injector is a protocol-aware fault-injection middleware that interposes on the MCP tool-call path with controlled persistence and seedable schedules. Its fault model is calibrated to the incident distribution of the public MCP ecosystem, mined from 57,240 issue and pull-request threads across 1,774 open-source MCP repositories. 19
Comparison: A 1,350-cell experiment crossing two backbone sizes, 20B and 120B, three agent stacks — a custom multi-round plan-and-execute loop, the OpenAI Agents SDK and LangChain — and three fault conditions. 19
Results: Faults drawn from the permanent-heavy stress profile cut the task-completion score by 60%, from 8.08 to 3.23 out of 10. The 120B backbone retains no larger share of its fault-free score than the 20B (resilience-ratio model, p = 0.96), and its completion rate under faults is 91.1% against 95.6%, a 4.4-point difference with a 95% confidence interval of [−0.2, 9.3]. The three stacks differ in how well they absorb faults. 19
Resources: Code and data are at the project repository on GitLab. 19
Takeaway: In the two backbones tested, scale bought nothing in fault tolerance, and the agent stack accounted for more of the difference. A capability benchmark on healthy tools can therefore hide the fragility that a production deployment meets first. 19
Latent dynamics and forecasting
8. Paloma: Phase-Conditioned Residual Modulation for Time Series Forecasting
Area tag: Time-series forecasting
arXiv: 2610.25392, first submitted October 18, 2026. 8
Authors / institutions: Jingru Fei, Wei Fan, Kun Yi, Qi Zhang and Zhendong Niu, at the Beijing Institute of Technology, the University of Auckland, the State Information Center and Tongji University. 20
Status: Accepted by NeurIPS 2026 as a spotlight. 8
Problem: Forecasters that model periodicity explicitly subtract a periodic component and treat what remains as unstructured noise. The paper shows empirically and analytically that the residual still has structure, with deviations that depend on phase in both direction and magnitude. 20
Method: Two modules borrowed from phase and amplitude modulation in signal processing: a phase rotation module applying phase-conditioned rotations that capture directional change, and an amplitude scaling module applying phase-conditioned affine transforms that capture magnitude change. Both are packaged as a plugin that can be inserted into an existing forecasting backbone. 20
Comparison: Twelve datasets — ETTm1, ETTm2, ETTh1, ETTh2, Electricity, Traffic, Weather, Solar-Energy and PEMS03, 04, 07 and 08 — against DistDF and GTR from ICLR 2026, RAFT, FreDF, TimeXer, CycleNet, FilterNet, iTransformer, PatchTST and DLinear, averaged over four prediction horizons. 20
Results: Paloma is best on MSE for 7 of the 12 datasets and best on MAE for 11. ETTm1 MSE falls to 0.363 against 0.367 for GTR and 0.378 for DistDF. Attached as a plugin it improves a wide range of the baseline models. 20
Resources: Code is public at the project repository. 20
Loading content card…
Takeaway: For a forecasting team the plugin is the useful form, since it attaches to a backbone already in service. The table also marks its limit: on Traffic Paloma sits 0.047 MSE behind RAFT, so the gain depends on the dataset. 20
9. Meta-iLaD: Identifiable Latent Dynamics via Meta-Learning of Dynamics Environments
Area tag: Latent dynamics and identifiability
arXiv: 2610.27180, first submitted October 19, 2026. 9
Authors / institutions: Yubo Ye and Linwei Wang at the Rochester Institute of Technology, with Sweekar Piya and Xiajun Jiang at the Department of Computer Science, Rowan University. 21
Status: ICML 2026 proceedings paper; the arXiv record's journal reference gives PMLR 306:148560–148582. 9
Problem: Latent-dynamics models recover a latent state from high-dimensional time series, and the identifiability results behind them condition on predefined labels of the dynamics environment. A new environment then has no label to condition on, and the identifiability of the environment variable and of the dynamics themselves was left open. 21
Method: A feedforward meta-learner models a conditional prior over the environment variable and extracts it from a few-shot context. The paper establishes identifiability for the latent state, the environment variable and the dynamics jointly, without restricting the environment variable's dimension or the way it modulates the dynamics. 21
Comparison: TDRL and IDF as identifiable baselines that need predefined environment labels, plus ODE2VAE and MONODE, on MoCAP with held-out subjects for out-of-distribution evaluation and on Pendulum; metrics are reconstruction and prediction MSE over the first 75 and the next 50 time steps, with MCC scores for the recovered variables. 21
Results: On MoCAP out-of-distribution prediction, Meta-iLaD reaches an MSE of 78.7 against 108 for MONODE and 118 for IDF, and on out-of-distribution reconstruction 32.9 against 53.4 and 76.5. In-distribution prediction is 19.5 against 38.7 for MONODE and 32.1 for IDF. TDRL and IDF could not run the out-of-distribution setting at all, because it removes the labels they condition on. 21
Resources: Code is public at the project repository. 21
Loading content card…
Takeaway: The few-shot context is what replaces the missing environment label, and the identifiability result on the environment variable is what keeps prediction stable when the subject changes. The paper's evidence for both is concentrated on MoCAP and Pendulum, so a reader with a different dynamical system is extrapolating from two datasets. 21
Unlearning and its traces
10. TRACE: Data-Free Text Reconstruction Attacks against Approximate Unlearning in LLMs
Area tag: Machine unlearning and privacy
arXiv: 2610.31995, first submitted October 22, 2026. 10
Authors / institutions: Mengying Zhang, Derui Wang, Ibrahim Khalil, Bowen Liu, Xiaoyu Xia and Minhui Xue, across RMIT University, CSIRO, Adelaide University and The Hong Kong University of Science and Technology. 22
Status: Accepted to the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). 10
Problem: Approximate unlearning is the affordable alternative to retraining, and deployments often keep the pre- and post-unlearning snapshots for auditing and rollback. The two snapshots together carry a side channel: the parameter difference can hold traces of the text that was supposed to be gone. 22
Method: Given both snapshots and no data-specific prior knowledge, TRACE first narrows the search space to a compact set of candidate tokens drawn from embedding-level parameter differences. It then assembles sequences by contrastive scoring on the perplexity gap between the two models, and a two-stage explore-then-complete strategy recovers several distinct samples rather than one. 22
Comparison: A perplexity-only reconstruction baseline, across six unlearning methods — gradient ascent, gradient difference, KL minimization, NPO, NPO-Diff and NPO-KL — on LLaMA2-7B and 13B, Mistral-7B, Qwen2.5-3B, StableLM-3B-4E1T and TinyLlama-1.1B, over TOFU, MUSE and WMDP, scored by ROUGE-L recall and by average extraction success rate at ROUGE-L thresholds of 1.0 and 0.9. 22
Results: On TOFU with TinyLlama-1.1B after gradient-ascent unlearning, TRACE reaches 0.9084 ROUGE-L recall against 0.4383 for the perplexity-only method, and an exact-match extraction rate of 78.75% against 10.00%. After KL minimization on the same model the rates are 76.25% and 2.50%. On Qwen2.5-3B after gradient ascent, 67.50% against 0.00%. 22
Resources: Code is public at the project repository. 22
Loading content card…
Takeaway: The attack needs only the pair of snapshots, so the deployment decision it bears on is who can read them. The authors recommend managing snapshot access and defending against parameter-difference leakage, which sits outside what an unlearning algorithm itself controls. 22
World models
11. AEON: Unifying Video and 3D World Models
Area tag: World models and 4D generation
arXiv: 2610.28631, first submitted October 20, 2026. 11
Authors / institutions: Weiqi Zhang, Wenyuan Zhang, Yu-Shen Liu and Junsheng Zhou, at the School of Software, Tsinghua University. 23
Status: Accepted to NeurIPS 2026. 11
Problem: World models split along two routes. Video generators such as Sora produce realistic appearance and lose 3D consistency; 3D generators such as Marble produce consistent geometry and give static, less realistic observations. 23
Method: A generative model sits in the text- or image-conditioned video latent space and emits 4D Gaussian Splatting as the explicit world representation. A 4D reconstruction model trained to predict splats and camera poses from monocular video initializes the decoder, and a domain adapter distills generative video latents into reconstruction latents through the video VAE. 23
Comparison: 4D generation against 4DNex and Lyra on 30 image-text scenes, video-to-4D reconstruction against L4GM, 4DGT, MoVies, NeoVerse and StreamSplat on Dycheck and DaVis, and novel-trajectory rendering against TrajectoryCrafter, RecamMaster and NeoVerse. 23
Results: In 4D generation with the Wan 2.2 backbone, AEON scores 0.913 on I2V subject consistency, 0.875 on subject consistency and 0.531 on aesthetic quality, against 0.865, 0.864 and 0.459 for Lyra and 0.796, 0.790 and 0.342 for 4DNex. On DaVis reconstruction PSNR is 24.661 against 23.929 for NeoVerse, with LPIPS 0.243 against 0.273; on Dycheck 14.487 against 14.269. On novel trajectories it takes the best score on four of six VBench metrics, with aesthetic quality 69.02 against 64.81. 23
Resources: Code and videos are at the project page, weiqi-zhang.github.io/AEON. 23
Takeaway: At inference the model needs neither camera poses nor depth, and only the first training stage needs annotated video, so later stages can use raw footage. The reconstruction margin over NeoVerse is 0.73 dB on DaVis and 0.22 dB on Dycheck, which puts it close to its own 4D reconstruction teacher at 14.413 dB on Dycheck. 23
Clinical imaging
12. TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation
Area tag: Medical imaging and report generation
arXiv: 2610.26271, first submitted October 19, 2026. 12
Authors / institutions: Yiheng Dong, Yi Lin, Shilong Huang, Xiyan Yang and Xin Yang, at the Huazhong University of Science and Technology and at HKUST. 24
Status: Accepted by CVPR 2026. 12
Problem: Radiology report generation is mostly single-timepoint, so it cannot describe how a finding evolved. Longitudinal methods add prior studies but merge images from different visits into one representation space, which blurs disease-specific semantics and weakens temporal reasoning. 24
Method: Stage I learns a temporal-decoupled representation that separates the evolution pattern from disease-specific features and writes reports for both the prior and the current study. Stage II looks for diagnostic inconsistencies in the prior report and iteratively corrects both reports through error-sensitive feedback, alternating the two stages for three iterations. 24
Comparison: Longitudinal-MIMIC, against single-image methods (R2Gen, R2GenCMN, R2GenGPT, PromptMRG, EKAGen, GMoD, RADAR) and longitudinal ones (Prefilling, HERGen, STREAM, MLRG, LLM-RG4, HC-LLM, Diff-RRG), scored on BLEU-n, METEOR and ROUGE-L plus clinical-efficacy precision, recall and F1 over 14 disease labels from CheXbert. 24
Results: Stage II reaches BLEU-4 0.124, METEOR 0.185, ROUGE-L 0.287 and clinical F1 0.511. The strongest prior longitudinal method, Diff-RRG, reaches 0.120, 0.164, 0.276 and 0.474. Stage I alone already gives F1 0.483, and adding its temporal branch to the baseline lifts F1 from 0.404 to 0.422. 24
Resources: Code is public at the project repository. 24
Loading content card…
Takeaway: Most of the clinical gain comes from separating evolution from disease features, and the refinement stage adds the rest by fixing inconsistencies in the prior report it is given. The evaluation covers one dataset, Longitudinal-MIMIC, and the clinical metrics come from a label classifier over 14 findings. 24
What to open first
- Serving a low-bit model: BitMTP (2610.28500) if multi-token prediction is on your roadmap and you want one backbone instead of a drafter, and LEMS (2610.32878) if you compress by SVD and allocate rank by hand today. 1
- Paying for reasoning tokens: SteadyThought (2610.25045) for a preference objective at switch points, and Correct but Unselectable (2610.32235) before you add candidates to a reranking pool. 3
- Making an evaluation claim: Claims of emergence (2610.28479) if you are about to write that a system or a team is more than the sum of its parts. 5
- Training or shipping an agent: BIRD-RL (2610.29426) for stateful environments and the curriculum around them, and MCP-Injector (2610.32555) before you trust a tool-calling agent with production tools. 6
- Forecasting with a periodic backbone: Paloma (2610.25392) as a plugin that costs two modules. 8
- Modeling a system that shifts between environments: Meta-iLaD (2610.27180) for few-shot environment conditioning and the identifiability result behind it. 9
- Handling deletion requests or generating video worlds: TRACE (2610.31995) if snapshots of unlearned models are retained anywhere, and AEON (2610.28631) for a world representation that renders from new viewpoints. 10
- Writing clinical reports over time: TIM (2610.26271) for the two-stage decoupling and the refinement loop between prior and current reports. 12
References
- 1BitMTP arXiv record
arxiv.org
- 2LEMS arXiv record
arxiv.org
- 3SteadyThought arXiv record
arxiv.org
- 4
- 5Claims of emergence arXiv record
arxiv.org
- 6BIRD-RL arXiv record
arxiv.org
- 7MCP fault injection arXiv record
arxiv.org
- 8Paloma arXiv record
arxiv.org
- 9Meta-iLaD arXiv record
arxiv.org
- 10TRACE arXiv record
arxiv.org
- 11AEON arXiv record
arxiv.org
- 12TIM arXiv record
arxiv.org
- 13BitMTP paper
arxiv.org
- 14LEMS paper
arxiv.org
- 15SteadyThought paper
arxiv.org
- 16Correct but Unselectable paper
arxiv.org
- 17Claims of emergence paper
arxiv.org
- 18BIRD-RL paper
arxiv.org
- 19MCP fault injection paper
arxiv.org
- 20Paloma paper
arxiv.org
- 21Meta-iLaD paper
arxiv.org
- 22TRACE paper
arxiv.org
- 23AEON paper
arxiv.org
- 24TIM paper
arxiv.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.