Notable Papers from ICLR, ICML, NeurIPS, CVPR, EMNLP (2025–2026): An Award-Grounded Reading List

A reading list of the award-winning and outstanding papers across the major ML and CV/NLP venues for the 2025 cycle, plus the just-announced ICLR 2026 outstanding papers. Every paper title is linked to its arXiv preprint; the conference page is linked separately when useful.

As of today (2026-05-01), only ICLR 2026 has concluded for the 2026 cycle — CVPR 2026 (June), ICML 2026 (July), NeurIPS 2026 (December), and EMNLP 2026 (October, Budapest) have not yet happened. EMNLP 2026 submissions are due 2026-05-25.

Table of contents

ICLR 2025 — Outstanding Papers

ICLR 2025 selected three Outstanding Papers from 3,704 accepted submissions (award announcement).

• Safety Alignment Should Be Made More Than Just a Few Tokens Deep — Qi, Panda, Lyu, Ma, Roy, Beirami, Mittal, Henderson. Argues that current alignment is “shallow” — it only conditions the first few output tokens — and that this single weakness explains adversarial-suffix attacks, prefilling attacks, decoding-parameter attacks, and the easy reversal of safety via fine-tuning. Demonstrates a deeper-token training fix.
• Learning Dynamics of LLM Finetuning — Yi (Joshua) Ren, Danica J. Sutherland (UBC). Step-wise decomposition of how influence accumulates across responses during SFT and DPO. Identifies the “squeezing effect” that explains why running off-policy DPO too long pushes down even the desired outputs.
• AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models — Fang, Jiang, Wang, Ma, Shi, X. Wang, He, Chua. Projects parameter perturbations onto the null space of preserved knowledge, so edits don’t collide with what was already there. ~36.7% average gain on top of existing locating-then-editing methods, in a single line of code.

ICML 2025 — Outstanding Papers

Six Outstanding Papers in the main track, picked from ~3,200 accepted out of ~12,000 submissions (awards page).

• CollabLLM: From Passive Responders to Active Collaborators — S. Wu, Galley, Peng, Cheng, G. Li, Dou, Cai, Zou, Leskovec, Gao. Trains LLMs against multiturn-aware rewards estimated by collaborative simulation, instead of next-turn rewards alone. +18.5% task performance, +46.3% interactivity, and +17.6% user satisfaction in a 201-judge user study.
• Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions — J. Kim, Shah, Kontonis, Kakade, S. Chen. Theoretically and empirically, masked diffusion models train on intractable subproblems but at inference can pick the next token to decode. The “decode where most confident” heuristic alone takes Sudoku from 7% → ~90%.
• Roll the Dice & Look Before You Leap: Going Beyond the Creative Limits of Next-Token Prediction — Nagarajan, C. H. Wu, Ding, Raghunathan. Constructs minimal algorithmic creativity tasks; shows next-token learning is myopic, that teacherless and diffusion approaches are more creative, and that input-side seed-conditioning rivals output-temperature sampling.
• Conformal Prediction as Bayesian Quadrature — Snell, Griffiths. Reframes conformal prediction through a Bayesian lens; unifies Split CP and Conformal Risk Control as special cases of Bayesian quadrature, with richer uncertainty representation.
• Score Matching with Missing Data — Givens, S. Liu, Reeve. Adapts score matching to partial-coordinate missingness via two routes (importance-weighted and variational), with finite-sample bounds for the IW variant.
• The Value of Prediction in Identifying the Worst-Off — Fischer-Abaigar, Kern, Perdomo. Formal welfare analysis of using ML to find the most vulnerable individuals (rather than maximize aggregate outcomes) in social policy, grounded in a German long-term-unemployment case study. Comparing prediction against expanded bureaucratic capacity is the key contribution.

CVPR 2025

Both top awards went to 3D / inverse-rendering work (best papers page).

• VGGT: Visual Geometry Grounded Transformer — Wang, M. Chen, Karaev, Vedaldi, Rupprecht, Novotny (Oxford VGG / Meta AI). Best Paper. A single feed-forward Transformer that, in one second, jointly predicts camera parameters, depth maps, point maps, and 3D point tracks for one to hundreds of input views — beating optimization-based pipelines on most 3D benchmarks. (code)
• Neural Inverse Rendering from Propagating Light — Malik, Attal, Xie, O’Toole, Lindell (Toronto / Vector / CMU). Best Student Paper. First physically-based neural inverse rendering from multi-view time-resolved (transient) measurements, using a time-resolved extension of neural radiance caching. Enables decomposing captured light into direct/indirect components and time-resolved relighting.

EMNLP 2025 (Suzhou)

Single Best Paper plus 35 highlighted Outstanding Papers (awards page).

• Infini-gram mini: Exact n-gram Search at the Internet Scale with FM-Index — H. Xu, J. Liu, Choi, N. A. Smith, Hajishirzi (UW / AI2). Best Paper. FM-index–based system that makes 83 TB of text (Common Crawl Jan–Jul 2025, DCLM-baseline, Pile) exactly searchable by n-gram, with index size only 44% of the corpus. Practical infrastructure for contamination audits, membership inference, and grounding work.
• PAFT: Prompt-Agnostic Fine-Tuning — Wei, Y. Shu, Ou, Y. He, F. R. Yu. Outstanding Paper. Continually samples diverse synthetic prompts during SFT/RLFT so the model learns task-level structure rather than surface phrasing. +7% generalization to unseen prompts, 3.2× faster inference.
• Constructions are Revealed in Word Distributions — Rozner, Weissweiler, Mahowald, Shain. Outstanding Paper. Uses RoBERTa as a proxy for the language distribution and shows that constructions (in the construction-grammar sense) are visible as patterns of statistical affinity, including hard cases of superficially similar but semantically distinct constructions.
• To Mask or to Mirror: Human-AI Alignment in Collective Reasoning — C. Qian, Parisi, Bouleau, Tsai, Lebreton, Dixon (Google). Outstanding Paper. 748-participant “Lost at Sea” leader-election experiment, then matched LLM groups (Gemini 2.5, GPT-4.1, Claude Haiku 3.5, Gemma 3). Some models mirror human demographic biases; others mask and over-correct. Collective alignment is model-specific.

NeurIPS 2025

Four Best Papers and three Runner-Ups (award announcement).

Best Papers

• Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free — Qiu et al. (Alibaba Qwen). A learnable, head-specific sigmoid gate inserted right after Scaled Dot-Product Attention. Eliminates the attention-sink artifact (heads can now output zero), stabilizes training, tolerates higher learning rates, and improves scaling. Validated on 1.7B dense and 15B MoE up to 3.5T tokens.
• Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) — L. Jiang et al. Releases Infinity-Chat (26K real open-ended queries, 31K human annotations across ratings + pairwise preferences, 25 annotators per example) and a 6-category / 17-subcategory taxonomy. Documents both intra-model repetition and pronounced inter-model homogeneity across 70+ models.
• 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities — K. Wang et al. Most RL networks are 2–5 layers. Pushing contrastive RL to 1024 layers (with the right SSL classification objective rather than TD-regression) gives 2×–50× gains on locomotion and manipulation, in a goal-conditioned setting with no demonstrations or rewards.
• Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training — Bonnaire, Urfin, Biroli, Mézard. Identifies two distinct training timescales: τ_gen (when good samples appear) and τ_mem (when memorization sets in). τ_mem grows linearly with dataset size n while τ_gen stays constant, yielding a widening generalization window. Reframes early stopping as a structural necessity, not a heuristic.

Runner-Ups

• Optimal Mistake Bounds for Transductive Online Learning — Z. Chase, Hanneke, Moran, Shafer. Closes a 30-year-old open problem: the optimal transductive mistake bound is Θ(√d) where d is Littlestone dimension — exponentially tighter than prior Ω(log log d) / Ω(√log d) / Ω(log d) bounds. Establishes a quadratic gap between transductive and standard online learning.
• Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? — Y. Yue et al. (Tsinghua LeapLab). RLVR-trained LLMs beat their base models at small k in pass@k, but the base models win at large k. The reasoning paths RLVR produces are already in the base model’s sampling distribution — so the trained capability boundary actually narrows. Distillation, by contrast, genuinely expands it.
• Superposition Yields Robust Neural Scaling — Y. Liu, Z. Liu, Gore. First-principles derivation of neural scaling laws from representation superposition: in the strong-superposition regime (more features than dimensions), loss scales as L ∝ 1/m with model width. Validated on OPT, Pythia, Qwen via a sparse-autoencoder framework. Distribution-robust unlike previous explanations.

ICLR 2026 — Outstanding Papers (just announced)

Two Outstanding Papers and one Honorable Mention from 5,355 accepted of 19,525 submissions (27.4% acceptance) (award announcement).

• Transformers are Inherently Succinct — Bergsträßer, Cotterell, Lin. Proves Transformers are doubly exponentially more succinct than finite automata, and exponentially more succinct than RNNs and LTL, when representing formal languages. As a corollary, verifying properties of Transformers is EXPSPACE-complete — formally intractable.
• LLMs Get Lost in Multi-Turn Conversation — Laban, Hayashi, Y. Zhou, Neville. Across 200,000+ simulated conversations on six generation tasks, every top open- and closed-weight LLM tested shows an average 39% drop going from single-turn to multi-turn underspecified instructions. Decomposes the drop into a small aptitude loss and a large unreliability increase.

What’s still to come in 2026

• CVPR 2026 — June 2026.
• ICML 2026 — July 2026.
• EMNLP 2026 — Oct 24–29, 2026, Budapest. Submission deadline May 25, 2026.
• NeurIPS 2026 — December 2026.

A few patterns worth noticing

A pattern across the 2025 best-paper slate: deflationary findings are getting awarded. “RLVR doesn’t expand reasoning beyond the base model” (NeurIPS runner-up), “LLMs lose 39% of their capability across multi-turn conversation” (ICLR 2026), “alignment is shallow and easily reversed” (ICLR 2025), “different LLMs collapse onto the same outputs” (NeurIPS 2025) — the field is rewarding work that punctures over-claims rather than pushing the next benchmark by 0.7%.

A second pattern: mechanistic explanations of empirical phenomena are back. Why diffusion models don’t memorize, why scaling laws hold (superposition), why attention sinks exist (gating), what creativity gaps next-token prediction has — these are theory papers grounded in clean experiments, not pure benchmark contests.

A third: in 3D vision, the feed-forward Transformer has eaten the optimization-based pipeline. VGGT does in one second what bundle-adjustment-style pipelines do in minutes, and wins on accuracy.

Sources for everything above are linked inline; the conference award pages are the canonical lists if you want the full slate of Outstanding/Honorable Mention papers I didn’t pull out individually.