AI Research in 2025: The Year Reasoning, Agents and Test-Time Compute Broke Out
2025 was the year the field's focus shifted from building bigger models to making models reason longer, interact, improve themselves and help with discovery. Reinforcement learning produced reasoning. Thinking time became something you could scale. Models wrote their own practice problems. Agents rewrote their own code, and AI started discovering algorithms. This guide covers eighteen papers that set up 2026's self-improving-agent era. They are grouped by the research direction they established rather than ranked. Each gets a plain-English explanation and an animation, followed by what it found, why it matters and its sources.
01 The 2025 Awards at a Glance
| Conference | Top awards (covered here) | Other award winners |
|---|---|---|
| ICLR 2025 (April, Singapore) | Safety Alignment Should Be Made More Than Just a Few Tokens Deep · Learning Dynamics of LLM Finetuning · AlphaEdit; Honorable Mention: SAM 2 | Honorable Mentions: Data Shapley in One Training Run · Faster Cascades via Speculative Decoding |
| ICML 2025 (July, Vancouver) | CollabLLM · Train for the Worst, Plan for the Best | Roll the Dice & Look Before You Leap · Conformal Prediction as Bayesian Quadrature · Score Matching with Missing Data · The Value of Prediction in Identifying the Worst-Off |
| NeurIPS 2025 (December, San Diego) | Gated Attention for LLMs · Why Diffusion Models Don't Memorize · 1000 Layer Networks for Self-Supervised RL · Artificial Hivemind (Datasets & Benchmarks) | Runners-up include Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? |
| CVPR 2025 (June, Nashville) | VGGT | Best Student Paper: Neural Inverse Rendering from Propagating Light |
| ACL 2025 (July, Vienna) | Native Sparse Attention (reported) | — |
| ICCV 2025 (October, Hawaiʻi) | BrickGPT (Marr Prize) | Honorable Mentions: Spatially-Varying Autofocus · RayZer |
02 Paper Map
Twenty-two papers in nine research directions. Filter by award and click to jump:
03 🧠 Reasoning and Test-Time Compute
SFT-heavy reasoning → RL reasoning → inference-time scaling → self-generated curricula.
DeepSeek-R1: Incentivizing Reasoning via Reinforcement Learning
DeepSeek-AI · arXiv Jan 2025 · Nature, Sep 2025Instead of showing a model thousands of worked solutions, DeepSeek let it practise: it tries maths and coding problems, gets a point when the final answer checks out, and adjusts. With nothing else, the model taught itself to write longer chains of thought, double-check and backtrack — even pausing mid-solution to say, in effect, "wait, let me re-think this" (the paper's "aha moment").
- R1-Zero: pure RL from a base model raised AIME 2024 pass@1 from 15.6% to 71.0% (86.7% with majority voting)
- Emergent self-verification, reflection and longer reasoning
- R1 adds a small cold-start SFT stage and multi-stage post-training for readability
- Distilled Qwen and Llama models brought reasoning to the open ecosystem; the paper became the first major LLM paper peer-reviewed in Nature
It opened reasoning-model techniques to everyone and made RL the default route to reasoning.
s1: Simple Test-Time Scaling
Muennighoff et al. · arXiv Jan 2025 · EMNLP 2025Two simple tricks. First, fine-tune on just 1,000 carefully chosen reasoning examples. Second, control how long the model thinks: stop it when the budget is used up, or — when it tries to stop too early — append the word "Wait" so it keeps checking. More thinking time, better answers.
- s1-32B (Qwen2.5-32B fine-tuned on s1K) beats o1-preview on competition maths by up to 27% (MATH, AIME24)
- Budget forcing extrapolates: AIME24 50% → 57%
- Model, data and code fully open
Made test-time compute an open, reproducible research direction rather than a proprietary secret.
Absolute Zero: Reinforced Self-Play Reasoning with Zero Data
Zhao et al. · arXiv May 2025 · NeurIPS 2025What if a model wrote its own homework? One model plays two roles: a proposer that invents coding puzzles just hard enough to learn from, and a solver that tries them. A code executor checks both the puzzles and the answers, so no human data is needed.
- Trained with zero external data, the Absolute Zero Reasoner reaches state-of-the-art among zero-setting models on coding and maths
- Outperforms models trained on tens of thousands of human-curated examples
- Works across model sizes and families
An early example of a system generating the experiences it learns from — a building block for self-improvement.
04 🤝 Agents, Interaction and Early Self-Improvement
From answering one turn well to collaborating, learning from interaction and rewriting one's own code.
Darwin Gödel Machine
Zhang, Hu, Lu, Lange & Clune · arXiv May 2025 · ICLR 2026A coding agent that edits its own code — its tools, prompts and workflow — then tests the new version. Every working variant is kept in a family tree, so later generations can branch from any ancestor, like evolution.
- SWE-bench 20.0% → 50.0%; Polyglot 14.2% → 30.7%
- Discovered improvements (better editing tools, context management, peer review) transferred across models
- Also documented reward hacking — a warning for self-improving systems
One of the clearest bridges from agent engineering to recursive self-improvement.
CollabLLM: From Passive Responders to Active Collaborators
Wu et al. (Stanford, Microsoft) · ICML 2025Most chatbots are trained to give the best answer right now. CollabLLM trains them for the whole conversation: it simulates how the chat might continue and rewards replies that lead to a better outcome later — which often means asking a clarifying question first.
- Multiturn-aware rewards estimated by simulating future conversation turns
- Better task performance and interactivity; higher user satisfaction and less time spent in a user study
Anticipates 2026's focus on multi-turn reliability (see LLMs Get Lost in Multi-Turn Conversation).
RAGEN: Self-Evolution in LLM Agents via Multi-Turn RL
Wang et al. · arXiv Apr 2025Training an agent over many steps of interaction is unstable. RAGEN found a failure it calls the Echo Trap: the agent finds one pattern that earns some reward and repeats it, its answers stop varying, and learning stalls. Filtering for varied, informative attempts (StarPO-S) keeps it learning.
- StarPO: trajectory-level RL framework for multi-turn agents
- Echo Trap: reward-variance collapse, entropy drop, gradient spikes
- Stabilisers: variance-based trajectory filtering, critic baselines, decoupled clipping; diverse initial states help
A bridge from reasoning RL on single answers to agents that learn from environments.
05 🧱 Model Architecture and Attention
Attention architecture → understanding fine-tuning → targeted model editing.
Gated Attention for Large Language Models
Qwen team (Alibaba) et al. · NeurIPS 2025Attention decides how much each word looks at every other word. Models develop a strange habit: they dump lots of attention on the first token (an "attention sink") just to have somewhere to put it. Adding a tiny learned gate after attention lets each head say "nothing useful here" instead — cleaner, more stable and better at long texts.
- A head-specific sigmoid gate after scaled dot-product attention, tested across many large-scale runs
- Better performance, training stability and long-context extrapolation; removes attention-sink behaviour
- Adopted in Qwen3-Next
Shows the Transformer itself is still improving, not only the systems built around it.
Learning Dynamics of LLM Finetuning
Ren & Sutherland · ICLR 2025Fine-tuning is usually treated as a black box. This paper tracks, step by step, how learning one example shifts the model's answers to other examples. It explains odd effects — like preference training (DPO) run too long making even the good answers less likely, because probability gets "squeezed" into one favourite response.
- A framework decomposing how each update influences predictions on other inputs
- Explains hallucination patterns after SFT and the "squeezing effect" in off-policy DPO
- Suggests simple fixes, e.g. training on rejected responses first
Turns fine-tuning from trial and error into something that can be predicted.
AlphaEdit: Null-Space Constrained Knowledge Editing
Fang et al. · ICLR 2025Suppose a fact changes — a country gets a new leader. Editing the model's weights to learn it often breaks other facts nearby. AlphaEdit only changes the weights in "directions" that the preserved knowledge doesn't use (its null space), so the new fact goes in and everything else stays put.
- Projects each edit onto the null space of preserved knowledge before applying it
- Average 36.7% improvement over the best baseline — by adding one line of code to MEMIT
Precise, controllable updates to what a model knows — important for trustworthy models.
06 🌫️ Diffusion and Alternatives to Autoregression
Is next-token prediction the only path forward?
Train for the Worst, Plan for the Best: Token Ordering in Masked Diffusions
Kim, Shah, Kontonis, Kakade & Chen · ICML 2025Masked diffusion models fill in a masked sequence in any order. Training is hard because they must handle every order, including terrible ones. But at inference time you can choose the order: fill in the cells you're most sure of first — like solving a Sudoku by starting with the obvious squares.
- Training on all orders includes computationally hard sub-problems
- Adaptive, confidence-based decoding needs no retraining: Sudoku accuracy 6.88% → 89.49%, beating autoregressive models trained with the correct order
Shows when flexible ordering is an advantage — the 2025 counterpart to 2026's Flexibility Trap.
Why Diffusion Models Don't Memorize
NeurIPS 2025Big image models could simply memorise their training pictures. Why don't they? Training has two clocks: one when the model starts producing good new images, and a later one when it starts copying training images. The second clock slows down as the dataset grows — leaving a wide window where you stop training with a model that generalises.
- Generalisation time stays roughly constant; memorisation time grows linearly with training-set size
- Implicit dynamical regularisation in training explains why over-parameterised diffusion models generalise
Theoretical grounding for training generative models that create rather than copy.
07 🛡️ Safety, Reliability and Evaluation
Can models reason? Are they robust? Aligned deeply? Becoming too similar?
Safety Alignment Should Be Made More Than Just a Few Tokens Deep
Qi et al. · ICLR 2025A safety-trained model mostly learns to start its answer with "I can't help with that." The paper shows the safety lives almost entirely in those first few words. Trick the model past them — by pre-filling "Sure, here's how…" — and it happily continues. The fix: train it to recover and refuse later in the answer too.
- Alignment mostly changes the distribution of the first few output tokens
- Explains prefilling attacks, some jailbreaks and fine-tuning attacks
- Remedies: data augmentation with recovery examples; a token-wise constrained fine-tuning objective protecting early tokens
A simple, unifying explanation for many jailbreaks — and a direction for deeper alignment.
The Illusion of Thinking
Shojaee et al. (Apple) · NeurIPS 2025Instead of benchmarks that may have leaked into training, the authors used puzzles with a difficulty dial, like Tower of Hanoi with more and more disks. Reasoning models beat ordinary ones at medium difficulty — but past a point, accuracy collapses, and strangely they start thinking less, not more.
- Three regimes: standard models win at low complexity, reasoning models at medium, both collapse at high
- Reasoning effort rises with complexity, then declines despite spare token budget
- Contested: follow-up work argued some failures reflect output-length limits and evaluation choices
Pushed evaluation from final-answer accuracy to how reasoning behaves as problems scale.
Artificial Hivemind: The Open-Ended Homogeneity of Language Models
Jiang et al. (UW, AI2 et al.) · NeurIPS 2025 Datasets & BenchmarksAsk 70 different AI models to "write a metaphor about time" and you'd expect 70 different ideas. Instead most say nearly the same thing ("time is a river…"). Each model repeats itself, and different models from different companies converge on the same answers.
- Infinity-Chat: 26K real-world open-ended queries with 31K dense human annotations
- Strong intra-model repetition and inter-model convergence across 70+ LLMs
- Reward models and LLM judges are poorly calibrated to diverse human preferences
Raises concerns about diversity, homogenisation of human thought and preference optimisation.
08 🔬 AI Doing R&D
AI → experiment → objective feedback → improved algorithm → next experiment.
AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery
Google DeepMind · May 2025 (white paper, arXiv Jun 2025)Gemini writes many candidate programs; automatic tests score each one; the best are kept and mutated again — evolution for code. Because every program is scored objectively, it can discover genuinely new algorithms, including ones that speed up the training of the very model that runs it.
- 4×4 complex matrix multiplication with 48 scalar multiplications, improving on Strassen's 1969 result
- Data-centre scheduling heuristic recovering ~0.7% of Google's stranded compute
- 23% faster Gemini matrix-multiplication kernel → ~1% less Gemini training time; hardware-design and FlashAttention improvements
The clearest 2025 precursor to AI improving the process that builds AI.
The AI Scientist-v2: Workshop-Level Automated Discovery
Sakana AI · arXiv Apr 2025An AI that does the whole research job: comes up with an idea, designs and runs experiments by exploring a tree of options, analyses results, draws figures and writes the paper. One of its fully AI-written papers scored above the acceptance threshold at an ICLR workshop.
- Agentic tree search over experiments; vision-language feedback on figures
- 3 papers submitted to an ICLR 2025 workshop; 1 averaged 6.33 — above the acceptance threshold (withdrawn by agreement)
- System later published in Nature (March 2026)
Foreshadowed 2026's automated-R&D wave (see AI Doing R&D).
09 🤖 Reinforcement Learning Beyond LLMs
The scaling story isn't unique to language models.
1000 Layer Networks for Self-Supervised RL
Wang et al. · NeurIPS 2025Reinforcement-learning agents usually use shallow networks (2–5 layers). Here a robot learns to reach goals with no reward and no demonstrations, and the authors made its network up to about 1,000 layers deep. Deeper networks unlocked new skills — walking and manipulating far better.
- Scaled self-supervised (contrastive) goal-conditioned RL networks to 1,024 layers
- 2× to 50× performance gains on locomotion and manipulation; qualitatively new behaviours emerge with depth
Suggests the scaling story isn't unique to language models.
10 👁️ Multimodal Foundation Models
Foundation-model ideas spreading from language into vision and video.
SAM 2: Segment Anything in Images and Videos
Ravi et al. (Meta FAIR) · ICLR 2025Click on an object once, and SAM 2 outlines it — in a photo, and then frame after frame through a video, even when it moves or is briefly hidden, thanks to a memory of what it has seen.
- Unified promptable segmentation for images and video with a streaming memory
- Real-time video segmentation; large SA-V video dataset released
Foundation-model ideas extending from language into persistent visual understanding.
11 🆕 Also Awarded in 2025: CVPR, ACL, ICCV and NeurIPS
Added 28 October 2026: award winners this article originally missed.
VGGT: Visual Geometry Grounded Transformer
Wang, Chen, Karaev, Vedaldi, Rupprecht & Novotny (Oxford VGG, Meta AI) · CVPR 2025Turning a pile of photos into a 3D model used to take a long, fiddly optimisation: guess where each camera was, match points between photos, refine, repeat. VGGT does it in one pass: show it one photo or hundreds, and a single Transformer returns the camera positions, a depth map for each photo and the 3D points, in about a second.
- A feed-forward network predicts camera parameters, point maps, depth maps and 3D point tracks from one to hundreds of views
- Runs in under a second and needs no slow geometry post-processing for strong results
- Code and models released (the original checkpoint is for non-commercial use)
3D reconstruction becomes a foundation-model problem: one network, no hand-built pipeline. It sets up the 2026 CVPR winner D4RT, which does the same for moving scenes.
Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
Yuan et al. (DeepSeek-AI, Peking University) · ACL 2025Normal attention makes every word look at every earlier word, so long documents get very expensive. NSA reads a long text the way you skim a book: a compressed summary of everything, a few chosen passages in full detail, and the nearby text just before. Unlike earlier tricks bolted on after training, the model is trained this way from the start, and the design matches how GPUs actually move data.
- Three attention branches: coarse token compression, fine-grained token selection and a sliding window
- Models pretrained with NSA match or beat full attention on general, long-context and reasoning benchmarks
- Reported speed-ups over FlashAttention at 64K tokens of roughly 9× (forward), 6× (backward) and 11.6× (decoding), as cited in secondary summaries
Long context gets cheaper without giving up quality, and sparse attention becomes something you design into pretraining rather than add later. It sits between 2019's Sparse Transformer and the 2026 proof of why local attention helps.
Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Yue et al. (Tsinghua LeapLab) · NeurIPS 2025After DeepSeek-R1, it looked as if RL taught models new ways to reason. This paper checks by letting each model try a problem many times (pass@k: is any of k attempts right?). With one try, the RL-trained model wins. With hundreds of tries, the original model solves more distinct problems. RL mostly teaches the model to pick answers it could already find, faster, rather than to find new ones.
- Across maths, coding and visual reasoning, RL with verifiable rewards raises pass@1, but at large k base models solve more unique problems
- RL amplifies rewarded solution paths and narrows exploration, so it stays within the base model's limits
- Distillation from a stronger model, by contrast, can add genuinely new reasoning patterns (as summarised in coverage of the paper)
It reframes the R1 story: RL makes models more reliable, but pushing the frontier of what they can solve may need new data, distillation or better exploration.
Generating Physically Stable and Buildable Brick Structures from Text (BrickGPT)
Pun, Deng, Liu, Ramanan, Liu & Zhu (Carnegie Mellon) · ICCV 2025Ask for "a small red chair" and BrickGPT designs it brick by brick, like a language model writing word by word. The twist is physics: after each brick it checks whether the structure would stand up, and if not it rolls back and tries again. The results can be built by hand or by a robot arm.
- An autoregressive language model predicts the next brick; invalid or unstable steps are pruned with a validity check and physics-aware rollback
- StableText2Brick dataset: over 47,000 stable brick structures of more than 28,000 objects, with captions
- Designs were assembled by people and by robotic arms
Generation that must obey the physical world, not just look right: a pattern that matters for robotics, design and manufacturing.
Also from 2025, not yet confirmed on the conferences' own pages: RLC 2025 outstanding papers reportedly included Deep RL with Gradient Eligibility Traces (Elelimy et al., theory; the forerunner of 2026's Gradient Iterated TD) and PufferLib 2.0 (RL at a million steps per second); COLM 2025 reportedly recognised The Implicit Curriculum Hypothesis.
12 Test-Time Compute, Simulated
Two 2025 papers pull in opposite directions. s1 says more thinking helps; The Illusion of Thinking says reasoning collapses on hard enough problems. A toy model shows both can be true at once. Extra "Wait" re-checks catch errors while the problem fits the thinking budget, and nothing helps once it doesn't:
"""Test-time compute in a toy model: why "think longer" helps - until it doesn't.
A problem needs n reasoning steps; each step is wrong with probability 3%. The model has a budget of
160 step-tokens. After finishing, it may say "Wait" k times (s1-style budget forcing): each re-check
costs n/2 tokens and finds and fixes one remaining error with probability 60%. If the solution plus
re-checks do not fit in the budget, the model runs out and fails - a crude stand-in for the accuracy
collapse on very complex puzzles reported in "The Illusion of Thinking". Illustrative, stdlib only.
"""
import random
def solve(n, k, rng, err=0.03, budget=160, fix=0.6):
if n + k * n / 2 > budget: k = max(0, int((budget - n) // (n / 2))) # use only the re-checks that fit
if n > budget: return False
errors = sum(rng.random() < err for _ in range(n))
for _ in range(k):
if errors and rng.random() < fix: errors -= 1
return errors == 0
def acc(n, k, trials=20000):
rng = random.Random(n * 100 + k)
return sum(solve(n, k, rng) for _ in range(trials)) / trials
print("Accuracy for a 30-step problem as the model is told to 'Wait' and re-check k times")
for k in [0, 1, 2, 4, 8]:
print(f" k={k}: {acc(30, k):.1%}")
print("\nAccuracy as problems get longer (budget 160 step-tokens)")
print(f" {'steps':>5} {'no re-check':>12} {'up to 4 re-checks':>18}")
for n in [10, 30, 60, 100, 150, 200]:
print(f" {n:>5} {acc(n, 0):>12.1%} {acc(n, 4):>18.1%}")
print("\nExtra thinking fixes errors on problems that fit the budget; past it, accuracy collapses.")Output:
Accuracy for a 30-step problem as the model is told to 'Wait' and re-check k times
k=0: 40.1%
k=1: 61.9%
k=2: 77.1%
k=4: 92.8%
k=8: 99.2%
Accuracy as problems get longer (budget 160 step-tokens)
steps no re-check up to 4 re-checks
10 73.8% 98.6%
30 40.1% 92.8%
60 16.2% 65.7%
100 4.6% 13.7%
150 1.1% 1.0%
200 0.0% 0.0%
Extra thinking fixes errors on problems that fit the budget; past it, accuracy collapses.13 The 2025 → 2026 Transition
From bigger models to self-improving systems. Press play to walk the timeline:
14 Key Takeaways from 2025
A counterpoint worth knowing: a NeurIPS 2025 runner-up, Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?, found that RL with verifiable rewards mostly makes a model more reliable at reasoning it could already sometimes do, rather than creating genuinely new abilities. The debate over whether RL teaches or sharpens continues. It is covered in the RSI series' Data vs Reinforcement Learning.
Next: AI Research in 2024: The Year Context, Agents and World Models Took Shape →