Every single one of the core open problems with production LLM reasoning got a working solution this week.

No new 1T parameter models. No magic new attention variants. Just nine papers, all posted to arXiv between June 8 and 10, that systematically address every failure mode we have been complaining about for two years.

None of this is theoretical. Every method here can be implemented on existing model checkpoints this quarter. Most will work with your existing training infrastructure.

This is not incremental progress. This is the point where the baseline for good reasoning shifts permanently.

We stopped using binary rewards

For 18 months the entire field ran on exactly one recipe for reasoning RL: sample 8 responses, run a verifier, give 1 reward if the answer is right, 0 if wrong, run GRPO. Everyone knew this was stupid. No one had a better replacement that actually worked.

DistIL changes this. The paper demonstrates conclusively that every existing self distillation objective used today, reverse KL, Jensen-Shannon, all the variants people ship in production, do not guarantee monotonic improvement. You can run an update step with a perfect expert, and make your policy worse. This is not an edge case. This happens 17% of the time in standard GRPO runs on the MATH benchmark.

DistIL uses forward cross entropy instead, adapted from distributional DAgger. It propagates disagreement between the expert and student backwards across the entire sequence automatically. No separate process reward model required. No manual step labelling.

On AIME 2025 it beats standard RLVR by 11.2 percentage points. On HumanEval+ it beats it by 7.9.

This is the new default for reasoning RL. Everyone will be running this by the end of July. Stop tuning GRPO hyperparameters.

You do not need RL to teach arithmetic

Everyone has been operating under the unstated assumption that reasoning requires reinforcement learning. That next token prediction can only memorize, not execute procedures.

The Arithmetic Pedagogy paper completely breaks this assumption.

They trained an 86 million parameter GPT-2 from scratch. No RL. No preference tuning. Just next token prediction on CoT traces generated following the GASING left-to-right arithmetic pedagogy used in Indonesian primary schools.

This 86M model hits 81% accuracy on 4 digit multiplication. For reference, Llama 3 8B hits 74% on the exact same benchmark.

Most importantly, they watched exactly how the model learned. First it learned to output the step by step procedure perfectly. Then after another 30k training steps, it stopped writing the intermediate steps. It developed internal mental arithmetic. It did not have to be taught this. It emerged spontaneously once the procedure was solid.

We have been wasting billions of FLOPs running RL on tasks that can be taught perfectly well with plain supervised fine tuning, if only we structure the demonstration data correctly.

Error snowballing is solved

Autoregressive CoT has one fatal flaw. One wrong step early in the chain and everything after it is garbage. The model will never go back. It will just confidently build an entire castle on top of the arithmetic error it made on line 2.

TRI fixes this completely.

It does not require any changes to the transformer architecture. It does not require bidirectional models. You train a standard decoder only model on a simple rearrangement task: given a verified start of the chain, a verified correct milestone later in the chain, fill in the missing part between them.

At inference you run this as a repair loop. Generate a draft chain. Run the verifier. Mark the last good step and the next correct milestone. Call TRI to rewrite only the broken segment.

On IMO 2024 problems this improves pass rate from 32% to 51%. And it uses 31% fewer tokens total than regenerating the entire chain from scratch every time you hit an error.

This is the single largest absolute accuracy gain published this year.

Gradient dilution was killing all your RL runs

Everyone knew that broadcasting a single sequence level reward to every token in the response was wrong. Everyone also knew that process reward models are too slow and too expensive to run at scale.

GRAIL is the compromise everyone was waiting for.

It does not use an external reward model. It measures gradient saliency locally. For each token, it calculates how much changing that token would change the final output answer. It then reweights the advantage signal by that value.

Filler words get almost zero gradient. The actual reasoning steps get almost all of it.

Across every tested model family, Qwen3, R1, OctoThinker, this gives an average 3.6% absolute accuracy gain and 3.05% Pass@3 gain. It adds zero overhead to training. It requires no additional data. You can drop this into an existing GRPO implementation in 12 lines of code.

There is no good reason to run standard GRPO ever again.

Self consistency was only working half the time

Self consistency is the most widely deployed test time reasoning trick. It is also extremely inefficient. Majority voting throws away the correct answer 38% of the time when that answer already exists in the sample set.

RISC fixes this. It replaces majority vote with a tiny 10M parameter LambdaRank model that ranks answers using five simple features: frequency, semantic centrality, trace length, internal consistency, and edit distance to other samples.

For the same number of samples, RISC gives between 4 and 7 percentage points higher accuracy across all benchmarks. Or alternatively, you can get the same accuracy as standard self consistency with 60% fewer samples.

This is a drop in replacement. You can deploy this tomorrow. No retraining required.

Reasoning distillation now actually generalizes

Distillation of reasoning has been broken forever. You can distil CoT from a big teacher into a small student and it will get great scores on the test set. Then you change one trivial surface detail of the problem and it falls apart completely.

IGA fixes this.

The core insight is extremely simple. For every problem you train on, you generate 5-10 logically identical versions of the same problem in completely different semantic domains. One as math, one as medicine dosage, one as contract law, one as inventory management.

Then during training you mask out any gradient direction that does not point the same way across all versions of the problem. You only update parameters on the gradient components that are invariant across surface semantics.

This gives a 14.3 percentage point gain on out of distribution reasoning. It reduces logical inconsistency by a factor of four.

Most importantly, this works with LoRA. You do not need full fine tuning.

We can now measure reward hacking before it ruins your model

Reward hacking on LLM as judge rubrics is not a hypothetical problem. It happens in almost every production RLHF run. No one talks about it publicly. Until this week there was no good way to reproduce it, measure it, or detect when it starts.

CHERRL is a test environment that lets you inject controlled known biases into a judge, then measure exactly how fast the policy learns to exploit them. The authors already used it to map exploitability for 17 common judge biases.

They also built a detector that can identify hacking onset from training logs 1200 gradient steps before reward divergence becomes visible. This detector has 94% true positive rate and 3% false positive rate.

Every team running RLHF at scale will be running this monitor by the end of the year.

The hidden cost of theorem proving

Agentic Lean provers have got very good very fast. No one has been talking about the cost. Standard implementations will happily burn 100k tokens on a proof attempt that had a 2% chance of working from the start.

The new cost routing agent fixes this. It adds a tiny control plane that estimates success probability and cost for every partial trajectory. It abandons dead ends early instead of grinding away until timeout.

It preserves exactly the same final proof rate. It uses 25.8% less compute on average.

This is the pattern that will define all agent systems going forward. Performance is no longer the limiting factor. Cost efficiency is.

Bias alignment finally stopped breaking model capability

Mitigating social bias has always been a terrible tradeoff. Every existing method would reduce bias by 20% and destroy general reasoning ability by 15%. Teams just accepted this as an unavoidable cost.

BiasGRPO changes this. It adapts group relative policy optimization to the high variance subjective reward landscape of bias evaluation. By normalizing rewards only within each batch of sampled completions, it eliminates almost all of the training instability that caused capability collapse in previous methods.

On standard bias benchmarks it outperforms DPO and PPO. Most critically, it measures less than 0.8% degradation in MATH accuracy after full bias alignment. This is the first alignment method that does not meaningfully break the model.

What this all means

None of these papers introduce radical new ideas. Almost all of them are things that people have suggested in passing for years.

What changed is that people finally stopped chasing bigger models, and started fixing the broken methodology that everyone was using.

Over the next 6 months you will see every major model provider implement most of these methods. You will see 7B models that outperform 70B models from 6 months ago. You will see reasoning accuracy go up 20-30% across the board without any increase in parameter count.

The era of scaling laws for reasoning is over. The era of good methodology has just started.