AI research papers to read — the 2025-2026 reading list.

~105 papers, Jan 2025 → May 2026, sorted into the story they actually tell — not a flat date dump. 2025 was the reasoning year (RL teaches models to think). 2026 so far is the architecture & efficiency year (linear attention, deltas, Mamba, leaner model reports). Read the "Start here" eight, then dive into whatever theme you need.

I keep a running list of papers I want to read. It grew into a monster. So I cleaned it up and turned it into something useful: a map, not a pile.

The raw list was just dates and links. Useful for nobody. A date-ordered feed hides the most important thing — papers cluster into conversations. Five teams attack the same problem in the same month. One paper answers a question another one raised. If you read them in calendar order you feel the whirlwind but miss the argument. If you read them by theme, the field suddenly makes sense.

So that is what this is. Every paper from January 2025 to May 2026, grouped by what it is actually about, with a one-line "why read it" on each, a few longer notes on the landmarks, and reading roadmaps for different goals. Pick a lane and go.

How to use this list

Three ways, depending on who you are.

If you are new to the reasoning-LLM world: read the eight "Start here" picks below in order. They give you the vocabulary — chain-of-thought, RLVR, reward models, GRPO, test-time scaling — that every other paper assumes you already have. Skip everything else until those click.

If you work in the field and want to stay current: skim the theme tables, read the abstracts of anything in a theme you touch, and full-read two or three landmarks per theme. The "why read" column is written so you can triage in ten seconds.

If you are hunting for a research direction: read the last paper in each theme section. Those are the most recent, and the most recent paper in a theme usually states the open problems the earlier ones left behind.

The two eras at a glance

Before the tables, the big picture. Seventeen months split cleanly into two moods.

The hinge between them is efficiency. Once R1 proved you could train strong reasoning with reinforcement learning, 2025 became a year-long sprint to understand why it works, make it cheaper, port it to small models, and stop the model from over-thinking. By 2026 the frontier had shifted under the hood: people went back to the architecture itself — replacing or augmenting attention, reviving state-space models, and shipping leaner model reports where "active parameters" is the headline number.

Start here — the essential eight

If you read nothing else, read these, in this order. They are the load-bearing walls.

#	Paper	Why read
01 📍🔬	Meta Chain-of-Thought (System 2 Reasoning)	The conceptual frame for the whole year — models can learn the process of thinking, not just the answer. Read this first for the mental model.
02 📍🏛️	DeepSeek-R1	The paper that defined 2025. Pure RL (GRPO) on a base model → long CoT and "aha moments" with no supervised reasoning data. Everything downstream reacts to this.
03 📍🏛️	Kimi k1.5	The other January landmark — a different, equally serious RL recipe. Read alongside R1 to see what is essential vs incidental.
04 📍🔬	Demystifying Long Chain-of-Thought	The "what is actually going on" paper — how long CoT emerges, reward shaping, why models backtrack. Best explainer once you have seen R1.
05 📍🔬	Understanding R1-Zero-Like Training	The skeptic's counterweight — what zero-style training really teaches and where gains come from. Read right after R1 so you do not over-believe the hype.
06 📍🔬	Spurious Rewards: Rethinking Signals in RLVR	The plot twist — even random/wrong rewards can help, so RL often elicits latent ability rather than teaching it. Essential for reading RLVR results critically.
07 📍🛠️	LIMO: Less is More for Reasoning	Strong reasoning from a tiny, curated example set — the counter-narrative to "scale the data." Reframes how much data reasoning needs.
08 📍📦	Qwen3 Technical Report	The best end-to-end modern model report: hybrid thinking/non-thinking, the full post-training stack. Read last — it ties the threads together.

2025 H1 · Surveys & blueprints

Start a new area by reading a survey — it gives you the taxonomy and the citation graph for free. A cluster landed in the first weeks of 2025, which tells you the field knew something big was coming.

Date	Paper	Why read
Jan 8	📍🔬 Meta Chain-of-Thought (System 2 Reasoning)	The conceptual frame: learning how to think, not just the answer. Read first.
Jan 16	🔬 Towards Large Reasoning Models: A Survey	The map of "reinforced reasoning." Best citation graph to mine for further reading.

2025 H1 · The landmark reasoning models

The week of January 22 is when the year turned. Two heavyweight RL-for-reasoning recipes dropped on the same day, and OpenAI followed with a competitive-programming result that confirmed the direction at the absolute frontier.

Date	Paper	Why read
Jan 22	📍🏛️ Kimi k1.5	Serious RL recipe: long context, partial rollouts, length penalties. The "other" R1.
Jan 22	📍🏛️ DeepSeek-R1	The defining paper. Pure RL → emergent long CoT, open weights, GRPO. Non-negotiable read.
Feb 3	🏛️ Competitive Programming with Large Reasoning Models	OpenAI's frontier confirmation: scaled RL reasoning reaches elite competitive-coding levels. Where the ceiling is.

2025 H1 · Understanding long chain-of-thought

Once R1 worked, the obvious question was why. This cluster is the field trying to understand its own breakthrough — and it is the most intellectually rewarding section to read.

Date	Paper	Why read
Feb 5	📍🔬 Demystifying Long Chain-of-Thought Reasoning	How long CoT emerges, reward-shaping effects, why backtracking appears. Best explainer.
Feb 11	📍🔬 LLMs Can Easily Learn to Reason from Demonstrations — Structure, not content, matters	The killer finding: the shape of the reasoning trace teaches more than its correctness. Reframes data needs.

2025 H1 · Rule-based & zero-style RL (RLVR)

RLVR = reinforcement learning from verifiable rewards. No learned reward model, just a checker that says right/wrong (math answer, unit test, logic rule). This cluster is where the deepest debates of the year live — including the two "Start here" skeptics.

Date	Paper	Why read
Feb 10	🔬 Exploring the Limit of Outcome Reward for Math Reasoning	How far pure outcome rewards (no process supervision) can go. A lot, it turns out.
Feb 20	🛠️ Logic-RL	Rule-based rewards on logic puzzles — clean, reproducible RLVR you can actually run.
Mar 26	📍🔬 Understanding R1-Zero-Like Training: A Critical Perspective	The reality check on what zero-style RL really teaches. Read with R1.
Mar 31	🛠️📦 Open-Reasoner-Zero	Fully open R1-Zero-style recipe on a base model. The reproducible reference implementation.
Jun 12	📍🔬 Spurious Rewards: Rethinking Training Signals in RLVR	Random/wrong rewards still help some models → RL often elicits, not teaches. Field-shaking.
Jun 17	🔬 RLVR Implicitly Incentivizes Correct Reasoning in Base LLMs	The counter-counterpoint: argues RLVR really does sharpen correct reasoning paths. Read against Spurious Rewards.

2025 H1 · Reward models, critique & judges

RL needs a signal. When you cannot verify the answer with a rule, you need a model to judge — a reward model, a critic, or an LLM-as-judge. This cluster is about making that signal itself smarter, often by making the judge reason too.

Date	Paper	Why read
Feb 5	🛠️ Teaching Language Models to Critique via RL	Train a critic with RL so its feedback actually improves the generator. Self-correction, done properly.
Mar 31	🛠️ JudgeLRM: Large Reasoning Models as a Judge	Make the judge a reasoning model — it deliberates before scoring. Better, more reliable verdicts.
May 5	🛠️ RM-R1: Reward Modeling as Reasoning	Recasts reward modeling itself as a reasoning task. One of the cleaner ideas of the spring.
May 21	🛠️ RL Tango: Reinforcing Generator and Verifier Together	Co-train generator and verifier so neither overpowers the other. Stops reward hacking.

2025 H1 · Search & agentic reasoning

Reasoning gets much stronger when the model can act — query a search engine, retrieve documents, call a tool — mid-thought. This cluster trains that behaviour with RL instead of hand-written agent scaffolds.

Date	Paper	Why read
Mar 10	🛠️ R1-Searcher	RL teaches the model to decide when to search. End-to-end, no fixed pipeline.
Mar 12	🏛️🛠️ Search-R1	The widely-cited one: interleave reasoning and live search-engine calls, trained with RL. Read this for the agentic-RL pattern.
Mar 30	🛠️ RARE: Retrieval-Augmented Reasoning Modeling	Bakes retrieval into the reasoning model rather than bolting RAG on top. The deeper integration.

2025 H1 · Efficient reasoning & "when to think"

Reasoning models love to ramble — thousands of tokens for a question that needs ten. That is slow and expensive. This cluster teaches models to be concise, and even to decide whether a question needs deliberate thinking at all. By spring this was one of the hottest sub-areas.

Date	Paper	Why read
Feb 6	🛠️ Training Language Models to Reason Efficiently	Penalize length, keep accuracy. The opening shot of the efficiency wave.
Apr 7	🛠️ Concise Reasoning via Reinforcement Learning	RL objective that rewards short correct traces. Practical and easy to adopt.
May 19	📍🛠️ AdaptThink: Reasoning Models Can Learn When to Think	The model learns to skip the chain entirely on easy questions. Big inference savings.
May 19	🛠️ Thinkless: LLM Learns When to Think	Same instinct, different recipe — dropped the same day as AdaptThink. Read both; the convergence is the point.

2025 H1 · Less data & self-improving reasoning

How little supervision can you get away with? This cluster pushes toward zero — tiny datasets, unsupervised prefixes, self-play, no external rewards at all. It is where 2025 starts to feel genuinely strange.

Date	Paper	Why read
Feb 5	📍🛠️ LIMO: Less is More for Reasoning	Strong reasoning from a handful of curated examples. The data-efficiency banner.
Mar 4	🛠️ The First Few Tokens Are All You Need (Unsupervised Prefix FT)	Fine-tune on just the prefix of reasoning traces, unsupervised. Surprisingly effective.
Apr 11	🛠️ Genius: Purely Unsupervised Self-Training for Reasoning	No labels at all — the model bootstraps its own reasoning. Ambitious and a little wild.
Apr 29	🔬 RL for Reasoning with One Training Example	One example moves the needle a lot. The most extreme data-efficiency result of the year.
May 6	📍🏛️ Absolute Zero: Reinforced Self-play Reasoning with Zero Data	The model proposes its own tasks and solves them. No human data in the loop. Must-read.
May 26	🛠️ Learning to Reason without External Rewards	Internal signals (confidence, consistency) replace external rewards. Closes the loop with Absolute Zero.
May 29	🏛️ Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents	Agents that rewrite themselves and evolve. The most futuristic paper on the list — read it last in this cluster.

2025 H1 · Small reasoning models

Does the R1 magic survive at 1B–3B parameters and on a hobbyist budget? Mostly yes — and that democratization is one of the best stories of the year.

Date	Paper	Why read
Mar 20	📍🔬 RL for Reasoning in Small LLMs: What Works and What Doesn't	The honest practitioner's guide for small models. Read before you try it yourself.
Apr 22	🛠️ Tina: Tiny Reasoning Models via LoRA	Reasoning gains via cheap LoRA fine-tuning. You can reproduce this on one GPU.
Apr 30	📦 Phi-4-Mini-Reasoning	Microsoft pushing the limits of small math-reasoning models. Strong, well-documented recipe.
May 12	📦 MiMo: Unlocking Reasoning from Pretraining to Posttraining	End-to-end small-model story — what to do in pretraining so RL has more to work with.

2025 H1 · Domain & multimodal transfer

Does the reasoning recipe port out of math/code into finance, software engineering, vision, and "everything"? This cluster says yes, with caveats — and it is where the methods meet the real world.

Date	Paper	Why read
Feb 13	🛠️ Adapting Language-Specific LLMs to a Reasoning Model via Model Merging	Turn a non-English model into a reasoner in a day by merging. Cheap, practical, open recipe.
Feb 25	🏛️🛠️ SWE-RL: RL on Open Software Evolution	Meta trains reasoning for real software engineering using git history as signal. Important for code agents.
Mar 10	🛠️ LMM-R1: Strong Reasoning in 3B Multimodal Models	Two-stage rule-based RL brings reasoning to small vision-language models.
Apr 10	🛠️ VL-Rethinker: Self-Reflection in Vision-Language Models	Gets VLMs to reflect and self-correct with RL. The multimodal version of "thinking."
May 20	🛠️ General-Reasoner: Reasoning Across All Domains	Pushes RLVR beyond math/code to general domains. The "does it generalize" answer.

2025 H1 · Frontier reports & advanced RL

The big model reports and the algorithm papers that pushed RL further — off-policy guidance, prolonged training, entropy-aware updates, and a few genuinely new ideas (reinforcement pre-training, RL teachers, algorithmic discovery).

Date	Paper	Why read
Apr 21	🛠️ Learning to Reason under Off-Policy Guidance	Use off-policy data (e.g. a stronger model's traces) to guide RL. Sample-efficiency win.
May 2	📦 Llama-Nemotron: Efficient Reasoning Models	NVIDIA's efficient reasoning model family. Solid open report with a focus on inference cost.
May 12	🏛️📦 INTELLECT-2: Globally Decentralized RL	A reasoning model trained on volunteer compute across the internet. Wild systems achievement.
May 14	📍📦 Qwen3 Technical Report	The best all-round modern model report. Hybrid thinking/non-thinking, full post-training stack.
May 23	🛠️ QwenLong-L1: Long-Context Large Reasoning Models with RL	Extends RL reasoning to long context. The long-doc reasoning recipe.
May 26	🛠️ Enigmata: Scaling Logical Reasoning with Synthetic Puzzles	Generate verifiable puzzles at scale for RLVR. Solves the data-supply problem for logic.
May 30	🏛️🛠️ ProRL: Prolonged RL Expands Reasoning Boundaries	Train RL much longer than usual → genuinely new capabilities, not just sharpening. Important.
Jun 2	🔬🛠️ Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive RL	A few high-uncertainty tokens carry most of the learning. Beautiful, actionable analysis.
Jun 3	🛠️ Rewarding the Unlikely: Lifting GRPO Beyond Distribution Sharpening	Fixes a known GRPO failure (just sharpening the existing distribution). Direct sequel to the skeptic debate.
Jun 9	🏛️ Reinforcement Pre-Training	Brings RL into pretraining, not just post-training. A genuinely new framing — read it.
Jun 10	🛠️ Reinforcement Learning Teachers of Test-Time Scaling	Train "teacher" models that make test-time scaling work better. Clever indirection.
Jun 12	📦 Magistral	Mistral's reasoning model report. The European frontier entry — good comparison point to Qwen3.
Jun 16	🏛️ AlphaEvolve: A Coding Agent for Scientific & Algorithmic Discovery	DeepMind's agent that discovers new algorithms. A glimpse of AI-driven science. Showstopper.
Jun 23	🔬 Programming by Backprop: LLMs Acquire Reusable Algorithmic Abstractions	Code training teaches reusable abstractions, not just syntax. Explains why code helps reasoning.
Jun 26	🛠️ Bridging Offline and Online RL for LLMs	Unifies offline and online RL — the synthesis paper closing out H1 2025.

The bridge — June to December 2025

The back half of 2025 is the hinge of the whole story. This is where the reasoning era stopped being a novelty and turned into infrastructure, where the field's confidence got punctured and rebuilt, and where the efficiency-architecture thread that the 2026 section runs with matured — lightning attention (born in January, see MiniMax-01 below) scaled into a frontier reasoning model, sparse attention reached a flagship release, and recursive tiny models appeared.

Two things frame the half-year. First, the milestones: in July 2025 frontier reasoning models from OpenAI and Google reached gold-medal level at the International Math Olympiad with no tools — pure token reasoning — and repeated the feat at the ICPC programming contest in September. Reasoning had arrived at the top of human competition. Second, the backlash: Apple's "Illusion of Thinking" argued these models collapse past a complexity threshold, igniting the loudest methodology fight of the year. Progress and skepticism, at full volume, at the same time.

Reasoning — the skeptics and the tiny-model surprise

Date	Paper	Why read
Jun 7	🔬 The Illusion of Thinking (Apple)	Argues reasoning models collapse past a complexity threshold. The most-debated reasoning critique of the year — read it, then read the rebuttals.
Jun 27	🏛️ Hierarchical Reasoning Model (HRM)	27M-param brain-inspired recurrent model solves hard puzzles from 1000 samples, no CoT, no pretraining. Reasoning without a giant LLM.
Oct 6	🏛️🛠️ Less is More: Recursive Reasoning with Tiny Networks (TRM)	Samsung's 7M-param recursive model beats R1/o3-mini/Gemini-2.5-Pro on ARC-AGI. The "do you even need scale for reasoning?" bombshell.

Model releases & technical reports — the open frontier goes agentic

Date	Paper	Why read
Jun 16	🏛️📦 MiniMax-M1: Scaling Test-Time Compute with Lightning Attention	456B hybrid-MoE reasoning model, 1M-token context, lightning (linear) attention. Carries MiniMax's linear attention (see MiniMax-01, Jan 14) into a frontier reasoning model — the point where efficient attention meets RL reasoning.
Jul 28	📍🏛️📦 Kimi K2: Open Agentic Intelligence	1T-param MoE (32B active), MuonClip optimizer, large-scale agentic data + joint RL. The defining open agentic model of H2 — read the post-training section.
Aug 8	📍📦 gpt-oss-120b & gpt-oss-20b Model Card	OpenAI's first open-weight models since GPT-2. Apache-2.0 MoE. Historically and practically important — the closed lab opens up.
Aug 8	📦 GLM-4.5: Agentic, Reasoning & Coding (ARC) Foundation Models	355B MoE (32B active), hybrid thinking/non-thinking, strong agentic+SWE-bench scores. Best single window into the "ARC" framing that dominates 2026 reports.
Dec 1	📍🏛️📦 DeepSeek-V3.2: DeepSeek Sparse Attention	Fine-grained sparse attention drops attention cost from O(L²) to O(Lk) with near-zero quality loss. The cleanest handoff into the 2026 efficient-attention era.

Architecture & efficiency — the survey that names the shift

Date	Paper	Why read
Aug 13	📍🔬 Speed Always Wins: A Survey on Efficient Architectures for LLMs	The map of linear attention, sparse attention, SSMs, MoE, and hybrids — exactly the toolkit the 2026 papers build on. Read this to understand why the next section exists.

2026 H1 · The architecture & efficiency turn

Now the mood changes. The 2026 papers are much less about "teach the model to reason" and much more about "rethink the machine underneath." Attention gets challenged. State-space models come back. Model reports brag about active parameters and 8B-class frontier intelligence. If 2025 was about the training loop, 2026 is about the architecture.

2026 H1 · Linear attention, deltas & Mamba

The technical heart of 2026 so far. A cluster of papers attacks the quadratic cost of attention — through delta rules, residual reformulations, and a full state-space-model revival. This is the most exciting architecture work in years.

Date	Paper	Why read
Jan 1	🏛️🛠️ Deep Delta Learning	Opens 2026 with a new take on delta-rule learning. Sets the agenda for the year.
Mar 15	🛠️ Attention Residuals	Reformulates attention with residual structure for efficiency/quality. Foundational read.
Mar 16	📍🏛️ Mamba-3: Improved Sequence Modeling via State Space Principles	The state-space comeback paper. If you read one 2026 architecture paper, make it this.
Mar 31	🛠️ Attention to Mamba: Cross-Architecture Distillation	Distill a trained transformer into a Mamba model. The practical migration path.
May 13	🛠️ Delta Attention Residuals	Combines the delta and residual threads. Read after Attention Residuals + Deep Delta.
May 21	🛠️ Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention	Cleaner linear-attention memory by splitting erase/write. State of the art for the sub-area.

2026 H1 · Frontier model technical reports

The 2026 model reports share a theme: do more with fewer active parameters. Mixture-of-experts, hybrid Mamba-transformer stacks, and aggressive efficiency are the norm now, not the exception.

Date	Paper	Why read
Jan 6	📦 MiMo-V2-Flash Technical Report	Fast, efficiency-tuned successor to MiMo. Good look at a modern lean model stack.
Jan 13	📦 Ministral 3	Mistral's small-model line continues. Reliable reference for the small-but-capable tier.
Jan 29	🔬 Scaling Embeddings Outperforms Scaling Experts	Surprising scaling result: where to spend parameters. Will reshape model-design intuition.
Feb 4	📦 ERNIE 5.0 Technical Report	Baidu's frontier report. Important non-Western frontier comparison point.
Feb 11	📦 Step 3.5 Flash: Frontier Intelligence with 11B Active Parameters	Frontier-level quality at 11B active params. The efficiency headline of early 2026.
Feb 12	📦 Nanbeige4.1-3B: A Small General Model That Reasons, Aligns, and Acts	A 3B model that does it all. The small-model bar in 2026.
Feb 17	📦 GLM-5: From Vibe Coding to Agentic Engineering	Zhipu's report, explicitly aimed at agentic software engineering. Read for the agent angle.
Feb 18	📦 Arcee Trinity Large Technical Report	Another large open report — good for triangulating 2026 design conventions.
Apr 13	📦🛠️ Nemotron 3 Super: Hybrid Mamba-Transformer MoE for Agentic Reasoning	Hybrid Mamba-transformer + MoE in one production model. Where the architecture threads converge.
May 6	📦 ZAYA1-8B Technical Report	An 8B report worth reading for its training-detail transparency.
May 25	📍📦 MiniMax-M2: Mini Activations, Max Real-World Intelligence	The capstone of the 2026 efficiency story so far — tiny activations, big capability. Read last.

2026 H1 · Interpretability & the science of representations

Alongside the architecture push, a quieter but deep thread: understanding what is actually happening inside these models — massive activations, attention sinks, the geometry of representations, interpretable visual tokens. This is the "science of deep learning" maturing.

Date	Paper	Why read
Jan 30	🔬 LatentLens: Highly Interpretable Visual Tokens in LLMs	Makes vision tokens inside multimodal LLMs human-readable. Great interpretability work.
Feb 16	🔬 Symmetry in Language Statistics Shapes Model Geometry	Connects statistical structure of language to the geometry models learn. Deep and elegant.
Mar 4	📍🔬 The Spike, the Sparse and the Sink: Anatomy of Massive Activations & Attention Sinks	Definitive anatomy of two long-mysterious phenomena. The interpretability highlight of 2026 so far.
Mar 12	📦 Tiny Aya: Bridging Scale and Multilingual Depth	Small multilingual model done right. Read for the low-resource-language angle.

2026 H1 · One vision detour

The list is LLM-heavy, but one vision paper earns its place — a major new vision-transformer design that everyone working in multimodal will reach for.

Date	Paper	Why read
Feb 8	🏛️ ViT-5: Vision Transformers for the Mid-2020s	A from-scratch modern ViT redesign incorporating years of lessons. The new default vision backbone.

Beyond reasoning — the other essential threads (Jan 2025 – May 2026)

The reasoning-and-RL spine is the loudest story of the window, but it is not the only one. Read only those papers and you get a distorted map — you would think 2025 was only about teaching models to reason. It was not. Several other threads ran in parallel, and a few of them are exactly what the 2026 architecture section is built on.

The one worth flagging hardest: the efficient-attention revolution started in early 2025, not 2026. It is easy to assume the jump from "reasoning" to "Mamba and deltas" happened overnight, but lightning attention, native sparse attention, and test-time memory all shipped in January–March 2025. The 2026 papers are the second act of a play that opened a year earlier.

Efficient attention & new architectures (the 2026 prequel)

Date	Paper	Why read
Jan 1	📍🏛️ Titans: Learning to Memorize at Test Time	Google's neural long-term memory module — a test-time-updating memory beside attention, scaling past 2M tokens. The conceptual seed of the 2026 memory/delta work.
Jan 14	📍🏛️📦 MiniMax-01: Scaling Foundation Models with Lightning Attention	456B hybrid-MoE with linear "lightning" attention, 4M-token inference. The first frontier-scale proof that linear attention works. (M1, in the bridge, is its reasoning sequel.)
Feb 16	📍🏛️ Native Sparse Attention (DeepSeek NSA)	Hardware-aligned, natively trainable sparse attention that beats full attention. The direct ancestor of DeepSeek-V3.2's sparse attention. Must-read for the efficiency arc.
Mar 13	🔬🛠️ Transformers without Normalization (Dynamic Tanh)	Replace LayerNorm with a one-line tanh(αx) and match or beat it. Beautiful, surprising, and trivially adoptable.

Test-time scaling, done simply

Date	Paper	Why read
Jan 31	📍🏛️🛠️ s1: Simple Test-Time Scaling	1,000 examples + "budget forcing" (append "Wait" to make the model think longer) rivals o1-preview. The cleanest, most reproducible reasoning recipe of the year — start here for test-time scaling.

Diffusion language models — the other paradigm

Date	Paper	Why read
Feb 14	🏛️ LLaDA: Large Language Diffusion Models	An 8B diffusion LM (not autoregressive) competitive with LLaMA3-8B and beating GPT-4o on the reversal curse. The serious case that text need not be left-to-right.

Multimodal & vision-language

Date	Paper	Why read
Feb 19	📦 Qwen2.5-VL Technical Report	Flagship open VLM: dynamic resolution, precise grounding, hour-long video, document parsing. The open multimodal reference for 2025.
Mar 12	📦 Gemma 3 Technical Report	Google's 1–27B open multimodal family — long context via a clever local/global attention ratio, trained with distillation. The best-documented small-open report of the spring.

Agents that use a computer

Date	Paper	Why read
Jan 21	🏛️📦 UI-TARS: Pioneering Automated GUI Interaction with Native Agents	An end-to-end GUI agent that sees only screenshots and drives keyboard/mouse — beats GPT-4o-based scaffolds on OSWorld/AndroidWorld. The native computer-use direction in one paper.

Reading roadmaps by goal

Different goals want different paths through the list. Pick the one that matches you.

How to actually read a paper

A reading list is useless if every paper takes you three hours. Here is the triage method that lets you process this whole list without drowning.

Pass 1 — the 5-minute skim. Title, abstract, all figures and their captions, and the conclusion. For 70% of papers this is enough — you learn the claim and the headline result and move on. The "why read" column here is doing pass 1 for you.

Pass 2 — the 30-minute read. Only for papers in a theme you actually work in. Read the intro, the method section properly, and the main results table. Skip the related-work and most of the appendix. You should be able to explain the method to a colleague after this.

Pass 3 — the deep read. Reserve for the handful you will build on. Re-derive the key equation, check the ablations, read the appendix, and ideally find the code. For most people this is two or three papers a quarter, not twenty.

Patterns worth noticing across the whole list

Read ~105 papers as a set and some meta-patterns jump out — these are more useful than any single result.

1. Convergent discovery is the norm. AdaptThink and Thinkless dropped the same day with the same idea. R1 and Kimi did too. When multiple independent teams land on the same answer in the same week, that idea is real and you should trust it more than a lone result.

2. Every breakthrough triggers a skeptic wave. R1 → "Understanding R1-Zero" → "Spurious Rewards" → "RLVR Implicitly Incentivizes." The skeptics are not noise; they are how the field figures out what is actually true. Always read the critical follow-up before you build on a hot result.

3. The pendulum swings from data to compute to architecture. Early 2025: scale the RL. Mid 2025: actually you need almost no data (LIMO, one example). 2026: forget the training loop, change the machine. The frontier moves to wherever the current bottleneck is.

4. "Efficiency" beat "capability" as the headline. 2025 reports led with benchmark scores. 2026 reports lead with active parameters. The race quietly shifted from "smartest" to "smartest per FLOP."

References & extra reads

• DeepSeek-R1 · the load-bearing paper of the whole reasoning era
• Understanding R1-Zero-Like Training + Spurious Rewards · the two essential skeptic reads
• Mamba-3 · the architecture-turn anchor for 2026
• arXiv cs.CL recent · the firehose, for keeping this list alive after May 2026
• arXiv cs.LG recent · machine-learning firehose for the architecture papers

Takeaways

2025 taught models to think; 2026 is rebuilding the machine that thinks. That single sentence is the whole arc. RL-from-verifiable-rewards (R1, GRPO, RLVR) defined the reasoning year, then the field spent months arguing about whether RL teaches or merely unlocks reasoning — and that argument is still open. Now the energy has moved under the hood, to linear attention, delta rules, and the Mamba revival, with model reports that brag about active parameters instead of raw scores.

You do not need to read all ~105. Read the eight essentials, pick the roadmap that matches your goal, pass-1 everything in that theme, and pass-3 only what you will build on. The map matters more than the pile — that was the whole point of cleaning this up.

And keep it alive: arXiv does not stop in May 2026. When you find the next R1, slot it into the right theme, not the bottom of a date-sorted dump. That is how a reading list stays useful.

Curated reading list, Jan 2025 → May 2026. Dates are first arXiv submission; links go to abstract pages so you get versions, code, and BibTeX. Emoji tags (📍🏛️🔬🛠️📦) are my own triage marks, not official.