A skill layer could help agents coordinate mixed-media work

In one sentence

The authors ask whether giving a student model feedback on more of its response helps it learn. In their tests, feedback at places where the student and teacher split the text in exactly the same way worked better than adding the particular feedback they tested for places where the splits differ. The student first generated responses; the authors then compared its predictions with those of a fixed teacher. They tested both how many positions could receive feedback and how well different training choices performed on mathematics and coding questions.

Key concepts

• A tokenizer breaks text into pieces called tokens. One model might treat a word as one piece while another divides it into several, making their next-piece predictions difficult to compare directly.
• On-policy distillation trains a student on responses it generates itself, using a teacher model’s predictions as feedback at the points the student reaches.
• Strict alignment means one student token and one teacher token cover the same piece of text. Strict full compares their prediction probabilities over all shared tokens at those positions; Strict top-16 compares them only over the 16 shared tokens the student rates highest at each position.
• A mismatch group covers the same stretch of text with multiple tokens on at least one side. The authors tested additional span supervision there: feedback based on how likely each model was to produce its observed sequence of pieces.
• Coverage counts how much of a response can receive feedback. Supervision reliability asks whether that feedback helps learning; covering more positions does not, by itself, establish that it does.

A concrete example

Hypothetical illustration, not a study result: two models answer a programming question. Where both split a word into one identical piece, a trainer can compare their predictions for the next piece. Where one splits that word into several pieces, comparing the whole stretch requires a different rule. The paper tests whether adding one such rule helps the student.

What the researchers measured

Across Qwen2.5-7B-Instruct → Llama-3.2-3B-Instruct, Granite-4.1-8B → Phi-4-mini-instruct, and Granite-4.1-8B → Qwen2.5-7B-Base, the authors report that Strict top-16 retained at least 96% of each pair’s improvement from the undistilled student to Strict full in the full-average accuracy score. Strict full compares probabilities over all shared tokens; Strict top-16 uses only the student’s highest-rated 16 shared tokens at each strictly aligned position. The full-average score gives equal weight to average accuracy on the mathematics and code benchmarks; accuracy counts correct sampled answers under the paper’s scoring rules. For those same pairs, all 18 tested positive-weight settings that added the authors’ span feedback to strict feedback lowered full-average accuracy relative to the corresponding Strict full run. The authors also report that Strict full and Strict top-16 scored above the four evaluated alternative distillation methods on the full average for each pair. Their gradient measurements—a check of whether two training signals point toward similar model changes—showed weak or negative agreement at checkpoints from Strict full training. The authors say this may help explain the lower scores with added span feedback; it does not establish a cause.

Why it matters

Different tokenizers need not prevent useful teacher–student comparisons. For the model pairs and tasks studied, the authors report that many generated positions aligned exactly, and that adding their tested feedback for the remaining positions lowered the overall benchmark score. That separates the amount of available feedback from its measured learning value.

Where it might help

A possible use is choosing what teacher feedback to test when training models whose tokenizers differ. The paper reports benchmark comparisons, not a ready-to-deploy training recommendation.

Impact across sectors

• Possible impact for teams building coding assistants: the results could inform experiments on which teacher predictions to use during training; they do not establish gains in deployed assistants.
• Possible impact for educational-tool developers: the mathematics tests could inform research on smaller models that answer practice questions; the paper does not evaluate classroom use.
• Possible impact for model-training teams: the distinction between feedback coverage and learning value could guide comparisons of training objectives; it is not a measured reduction in training cost.

Where the evidence stops

The findings concern the named model pairs, training objectives and evaluations, not every way to handle mismatched tokens. The authors say they did no deduplication of their training-prompt pool. Their probability-mass analysis used student responses from before distillation and excluded responses that could not be scored under its rules; it was not a measurement of every generated position. The gradient analysis examined saved responses at checkpoints from training with only the strict loss, rather than directly measuring training updates in the added-span runs. The reported scores come from mathematics, code and additional ALFWorld benchmark evaluations, not deployments. General PtoP note: this arXiv source is a preprint; the supplied text does not establish peer review.

In one sentence

The study asks how to make a language model’s example-generating run cheaper without letting it drift too far from the run that learns from those examples. The authors’ answer is to let the first run guide how the second rounds its numbers. Imagine two people copying the same pattern with a small set of tiles: if a mark falls between tiles, each might choose a different one. The authors’ method, TRACE, records a compact clue about the tile chosen during generation and uses it during training. In the model, this is FP4 quantization: storing and calculating some values in a four-bit format. The authors apply it to selected parts of mixture-of-experts language models, which route work among specialised parts.

Key concepts

• A rollout is the model generating responses that will be used in reinforcement learning, where feedback helps update the model. Training then processes those responses to change the model.
• Quantization represents numbers with fewer possible values, like replacing a finely marked ruler with one that has fewer marks. FP4 is the four-bit format used for the low-precision rollout in the main tests.
• Train–rollout discrepancy is a difference between the training calculation and the calculation that generated its example. TRACE uses the rollout’s rounding outcome to guide the matching training calculation.
• Compact caching keeps only selected clues about rollout values, rather than full records. The default TRACE setting retains mantissa and scale information from the latter half of the model’s layers; these help reconstruct a reference, not recover every original value exactly.

A concrete example

Hypothetical illustration, not a reported test: a generated response leads two nearly identical calculations to opposite sides of a rounding boundary. Ordinary rounding picks different available values. TRACE would use a stored clue from generation to guide the corresponding training choice.

What the researchers measured

In the authors’ reasoning-task tests of Qwen3.5-35B-A3B, default TRACE with joint NVFP4 weight, activation and KV-cache rollout received an average reported benchmark score of 75.3, versus 74.9 for the BF16 rollout reference. That average combines scores from four named reasoning and coding benchmarks; it is not a success rate for everyday work. In a separate decoding-throughput test of the same model on four GB200 graphics processors, the authors report up to 5.4× higher throughput for TRACE than BF16 rollout at a 128K output length. Throughput here means generated tokens processed per unit time. The authors also report tests on three other named models, but these figures should not be transferred to them.

Why it matters

Generating many responses can make reinforcement-learning training costly. The paper distinguishes two goals that sound similar but are not: making each calculation individually close to a high-precision version, and making the generation and training calculations agree with each other.

Where it might help

A possible research use is designing lower-cost rollout generation when training a large language model with feedback. The paper tests model-training configurations, not a ready-to-use service or deployment.

Impact across sectors

• Possible impact for AI training teams: investigate whether rollout-guided rounding reduces generation costs in their own training setup; the paper does not establish their likely savings.
• Possible impact for coding-assistant researchers: explore the method in training experiments, since the authors tested models on coding tasks; this is not evidence of performance in a deployed coding assistant.
• Possible impact for teams studying long-running task assistants: examine low-precision training on their own tasks; the reported long-horizon result is a benchmark result, not a deployment.

Where the evidence stops

The authors say their rounding rule’s local guarantee does not guarantee that the entire model’s calculations or output behaviour will move steadily closer together. Compact caching reconstructs an approximate guide rather than the exact rollout value. TRACE also does not remove differences caused when training has moved on since an example was generated, known as policy staleness. In the stated default setup, only routed expert components and the KV cache use the low-precision rollout; other modules remain in BF16. The cited speed result is a particular hardware and output-length test. General PtoP note: this is an arXiv preprint, and benchmark results are not a tested deployment.

In one sentence

The study asks how an agent can continue to the next target when finishing one task has left it facing somewhere else. The authors trained Attacca on demonstrations that begin with searching, continue with approaching a target, and end with interacting with it. At each new task, the agent gets a picture showing what to seek, even though that picture comes from a different Minecraft world. During training, it also learns to predict whether the target appears in its current view and whether it is searching, approaching or interacting. Its actions use those predictions; it does not receive the correct answers during a run.

Key concepts

• State continuity means the next task starts where the previous one ended, with the agent’s position, view and changes to the world carried forward.
• A masked goal image is a reference picture that marks the wanted object. Attacca’s training pairs a demonstration with a picture of the same kind of object from another world, so matching the surrounding scenery is not a direct shortcut.
• Target grounding means deciding whether the wanted object is in the current view and where it is. During training, the authors provide a marked target area for each view, including an empty mark when the target is absent.
• Behavioral-phase conditioning means using a prediction of the current stage—Search, Approach or Interact—to guide action choices. The authors train that prediction alongside imitation of the players’ actions.

A concrete example

Hypothetical illustration, not a reported episode: after mining a log, an agent turns toward empty ground. Given a picture marking diamond ore in another world, it first looks for ore, then moves toward ore it sees, then mines it.

What the researchers measured

In the authors’ Mine single-task test, Attacca recorded 39.0% clean success across 200 episodes. Clean success means completing the requested interaction without interacting with the wrong class of object. The fine-tuned ROCKET-2 variant given the same different-world masked goal images recorded 0.235 clean success in that table; this is not the released ROCKET-2 checkpoint or the variant whose goal is built from its current view. In the Diamond Pickaxe chain, Attacca completed 54.0% of runs. That chain starts each stage from the state left by the preceding stage and used 50 paired policy seeds per method. These are Minecraft evaluation results, not tests of a deployed agent.

Why it matters

A plan can name the next job without placing its target in front of the agent. The authors focus on this gap between receiving a goal and being able to act on it.

Where it might help

The authors study low-level control in Minecraft, not a complete planning system. As a possible application, this way of training could inform an agent that must repeatedly find a new visual target after its surroundings change; use beyond Minecraft remains untested here.

Impact across sectors

• Possibility, not a proven deployment: game-agent developers could explore training controllers on the search that happens between scripted objectives.
• Possibility, not a proven deployment: robotics researchers could examine whether different-scene reference pictures help a controller find objects after a previous action changes its position. The paper does not test robots.

Where the evidence stops

The authors say the evaluation isolates the low-level controller. A scripted controller supplies each stage’s goal and detects completion; macros and helpers handle some crafting, cooking and item changes. Each long task chain uses one fixed world, and no method was trained on those chains. The authors have not evaluated Attacca connected to a planner or outside Minecraft. The training dataset also sets aside 10% of its episodes for validation; the single-task tests include both classes seen in training and classes held out from training. General PtoP note: this source is an arXiv preprint, not evidence of peer review.

In one sentence

The study asks whether an existing AI agent can handle work across several kinds of media without changing its underlying model. The authors’ approach is to give the agent a layer of reusable instructions, tool connections, and saved outputs. Imagine making a course from a recording: one step examines the recording, another prepares images, and a later step assembles material that draws on both. Omni-IO Skills represents such steps as tasks with stated dependencies, so independent work can run at the same time and later work can use saved files. This is a description of the authors’ system design; the measured evidence comes from their separate benchmark tests.

Key concepts

• A Skill is a reusable procedure that tells the agent when a task applies, what it needs, and what it should produce. The authors organize Skills for single operations, individual deliverables, and larger requests.
• The execution graph is a task plan that states which steps must finish before others begin. Steps without a dependency can be scheduled together.
• The Asset Registry is a record of supplied and produced files. It lets later steps refer to an output and lets a subsequent request refer to an earlier version.
• Tool and provider layers connect a task such as speech generation to an external service. The authors separate that connection from the Skill so the procedure need not name a particular service.

A concrete example

Hypothetical illustration, not a measured result: an event worker supplies a venue photo and asks for a poster and a short video. An agent could use the photo in both jobs, wait for any required source material before assembly, and retain the outputs for a later revision. The paper describes this kind of dependency and reuse; this particular request is not an additional test result.

What the researchers measured

The authors tested GPT-5.6 Sol and Claude Sonnet 5, each in its default environment and then with Omni-IO Skills added, on UniM-90, a fixed 90-instance subset of UniM. Input-support rate is the share of instances for which the agent can fully receive and process every input type. For GPT-5.6 Sol, the authors report 40.00% as a Base Agent and 100% with Omni-IO Skills. For Claude Sonnet 5, they report 38.89% and 100%, respectively. Their relative Semantic–Quality Coupled Score combines correctness and generation quality while reflecting performance across the full test set: it rises from 26.99 to 74.94 for GPT-5.6 Sol and from 27.82 to 77.78 for Claude Sonnet 5. The authors also report qualitative art-tutorial and product-promotion cases with both hosts; those cases are separate from the benchmark scores.

Why it matters

A working reader may need several outputs that share source material, rather than one answer in one format. The authors test a way to coordinate those outputs around an existing agent. General PtoP note: a benchmark result does not establish how a system will perform in a particular workplace.

Where it might help

The authors describe tasks involving documents, images, audio, video, code, and three-dimensional assets. Using the system for a particular organisation’s materials would be a possible application, not a deployment established by the benchmark.

Impact across sectors

• Possibility for education: a course producer could explore coordinating source recordings, illustrations, and teaching materials; the paper does not establish results for a live course-production service.
• Possibility for office work: a team could explore linking meeting audio to a report and presentation; this is not a measured workplace outcome.
• Possibility for marketing: a team could explore reusing product material across a poster and video; the paper’s product-promotion case is a qualitative demonstration, not evidence of campaign performance.

Where the evidence stops

The authors evaluate a fixed subset of UniM, not the whole benchmark. Each Base Agent retains its built-in tools and Skills, so the comparison is between the default environment and that environment with Omni-IO Skills added. The authors say the hosts’ default capabilities may differ and do not treat cross-agent scores as a model ranking. An absolute score for a Base Agent covers only the inputs it supports, whereas a relative score reflects the complete test set; those should not be read as the same setting. The case studies are qualitative. General PtoP note: the supplied source is an arXiv preprint, not evidence of peer review or a tested deployment.

In one sentence

A smaller math-trained model can help a larger model improve without supplying finished answers for it to copy. The authors tested this through on-policy distillation: the student generates its own answers, and a fixed teacher scores the student's choices as it writes. They began with Qwen2.5 Base models, gave them instruction-following training, and trained the teachers further on math problems using reinforcement learning, a method that updates a model using feedback on its answers. They then tracked students taught by smaller, equal-sized and larger teachers. Early in training, test accuracy rose in a fairly regular way as the student moved away from its starting behaviour. Later, improvement could slow, level off or reverse.

Key concepts

• On-policy distillation means the teacher supervises answers generated by the current student, rather than giving the student a fixed collection of teacher-written answers.
• Gold score is accuracy on the held-out math test set. It measures correct answers, not how closely a student imitates its teacher.
• Reverse KL divergence measures how far the student's word-choice probabilities have moved from its starting model. The authors used its square root as a measure of training progress, not as a count of training steps.
• A power law is a fitted relationship between quantities such as model size, teacher accuracy and a student's best observed accuracy. Here the authors fit separate laws for two distillation methods.
• Vanilla-OPD scores student-generated words using the fixed teacher's probabilities. Delta-OPD instead uses the change between the teacher before and after its math training.

A concrete example

Hypothetical illustration, not a paper result: a student starts solving a math problem in its own words. A smaller teacher provides feedback on the student's next-word choices; the student is not required to copy an answer written by the teacher. A separate set of problems checks whether its final answers become more accurate.

What the researchers measured

On the authors' Qwen2.5 math-reasoning experiments, all 25 Vanilla-OPD teacher–student runs initially improved in held-out accuracy. In the comparison of 17 shared teacher–student settings, Delta-OPD had a steeper early accuracy gain than Vanilla-OPD in 15; this slope compares gains at a matched amount of movement from each student's starting model. Delta-OPD also had a larger gain at its best observed checkpoint in 12 shared settings, mostly among weak-to-strong pairs. The authors report that their separate fitted peak-accuracy laws predicted held-out model scales, while the fitted early-gain rates were less precise. These are observations from training runs, not deployment results.

Why it matters

The authors report that teacher accuracy alone did not determine how well a student learned: in their fitted comparisons, a smaller teacher transferred better than a larger one at a matched teacher score. That makes teacher choice a question about the training relationship, not just which teacher gets more test answers right.

Where it might help

A possible use is planning comparisons between teacher sizes before committing to a math-model training run. The fitted relationships describe the authors' tested model family and setting; the paper does not establish that they will predict another family's results.

Impact across sectors

• Possible impact for AI model-development teams: compare smaller and larger math-trained teachers when planning student training, rather than choosing by teacher accuracy alone.
• Possible impact for education-tool developers: use the distinction between copying worked solutions and supervising a learner's own attempts to frame future research. The paper did not test an educational product.
• Possible impact for organisations building quantitative assistants: treat a model's best observed test accuracy and its later training behaviour as separate questions. The paper did not test workplace use.

Where the evidence stops

The authors limit their evidence to one pretrained family, one math-problem mixture, one run per teacher–student configuration, capped answer lengths and trajectories that were deliberately stopped. Teacher training and distillation used the same training prompts; accuracy was measured on a held-out test split. Evaluation records were aggregate rather than per problem, so the reported intervals do not measure variation across training seeds or resampled test problems. Some trajectories stopped before a departure from the early trend could be observed. The authors say transfer to other model families, tasks and training recipes is untested. They also warn that distillation can pass on errors, bias and unsafe patterns, and that held-out accuracy does not establish reliability or safety beyond this setting. General PtoP note: this arXiv source is a preprint, not a claim of peer review.

In one sentence

The study asks whether a language model trained from scratch with a relatively small budget can reach the benchmark-score range of open models trained with much larger budgets. The authors report that their model, HRM-Text, did so on most of the benchmarks they compared. They changed both how the model processes an instruction and what it learns to predict: it repeatedly revisits its internal working state, and it trains to produce the response rather than to reproduce the instruction. This is a result about the tested model and benchmarks, not a finding that every small training run will work the same way.

Key concepts

• Imagine training on a question and its answer. With the authors’ task-completion objective, the model is corrected for its answer, not for predicting the words of the question. HRM-Text was pretrained on instruction-response pairs rather than broad raw text.
• A recurrent model reuses parts of its computation, like revising a draft rather than starting with a new set of tools at each pass. HRM-Text has a fast-changing part for local revisions and a slower-changing part that carries context across cycles.
• PrefixLM is the authors’ attention pattern: the model can consider all words of an instruction together, but produces response words in order without seeing future response words.
• Repeated computation can make training signals unstable. MagicNorm is the authors’ method for keeping internal values bounded at each recurrent step. Their warmup deep credit assignment begins by sending training corrections through fewer steps, then extends how far back those corrections reach.

A concrete example

Hypothetical illustration, not a paper result: given the instruction “Explain how to sort these three receipts by date,” a task-focused model would be trained on its proposed explanation as the response. It would not receive the same training correction for reconstructing every word of the instruction.

What the researchers measured

The authors report that HRM-Text 1B was trained from scratch using 40 billion unique training tokens, with 60 billion tokens processed over the training run. On their evaluations, it scored 84.5% on GSM8K, a math-question benchmark, and 60.7% on MMLU, a broad knowledge-and-reasoning benchmark; these percentages are benchmark scores, not success rates in a workplace. The paper says its comparison with named open models used roughly 100–900 times fewer training tokens and 96–432 times less estimated training compute. In the authors’ matched-compute tests, changing the training objective, adding PrefixLM, and then using the HRM architecture each raised the reported benchmark scores in the tested configurations. Those are separate comparisons, not evidence that any one change alone produced the final model’s scores.

Why it matters

The authors’ comparisons address a practical research constraint: training a language model from scratch usually takes substantial data and computation. Their results suggest that model structure and the choice of what to predict can affect how much benchmark performance a training budget yields. The comparisons do not establish the same savings for other model sizes or real-world services.

Where it might help

The paper studies pretraining and benchmark evaluations, not workplace deployments. A possible use of its approach is to inform experiments in which a team compares model designs under a limited training budget.

Impact across sectors

• Research labs — possibility, not a demonstrated deployment: use the reported design and comparisons to plan smaller-budget pretraining experiments.
• Education technology — hypothetical possibility: investigate training on exercise-and-answer pairs; the paper does not test a classroom product.
• Customer-support software — hypothetical possibility: study whether response-focused training suits short request-and-reply tasks; the paper does not test a support service.

Where the evidence stops

The authors tested HRM-Text up to 1B parameters; whether the reported efficiency extends to larger HRM-Text models remains future work. They say recurrent steps also add computation when generating answers compared with a single-pass Transformer. Multi-turn use of PrefixLM needs careful handling of stored attention context. Their overlap check found a statistically significant possible contamination effect for HRM-Text 1B on DROP with one matching setting, but not with the other; the authors also report its score on a strictly clean DROP subset. Some comparison-model scores came from original papers rather than reruns. HRM-Text does not implement the external knowledge retrieval discussed as future work. General PtoP note: this arXiv source is a preprint, and benchmark results are not a deployment test.