The Next Frontier: AI That Learns from Experience

ROME, Sep 22, 2026

As the availability of new training data for language models plateaus, especially in translation, advancing performance requires exploring alternative methods and leveraging richer, non-textual contextual data.

Reasoning models offer a promising path forward. Unlike standard large language models that generate output directly, reasoning models utilize a step-by-step “chain-of-thought” process. They break down complex problems, evaluate options, verify intermediate results, and adjust course when needed before finalizing a response. Many of these models are specifically optimized for this workflow using reinforcement learning with expert human feedback, moving beyond simple step-by-step prompting.

Rather than relying exclusively on existing translations or external feedback, this is a process in which a model generates translations, evaluates multiple versions of its own output, and improves through reinforcement learning.

Key benefits

• Higher accuracy on multi-step problems — math, logic, coding, and multi-hop reasoning tasks see substantial gains because errors can be caught mid-process rather than compounding silently
• Better handling of ambiguity — the model can weigh competing interpretations explicitly instead of picking the first plausible one
• Improved reliability on tasks with verifiable answers — since the reasoning trace can be checked against a ground truth during training, models learn genuinely better problem-solving strategies, not just better-sounding output
• More transparent failure modes — when the model does get something wrong, the visible reasoning trace often shows where the breakdown happened, which is useful for debugging and building trust
• Reduced hallucination on certain tasks — because the model can verify sub-steps against each other before finalizing an answer

The tradeoff is cost and latency: reasoning models use more tokens and take longer per query. Reasoning models are a genuinely interesting fit for MT, though the picture is more mixed than it is for math or code, because translation quality is often subjective rather than verifiable.

Disambiguation. Many translation errors come from a source word or sentence being genuinely ambiguous without more context — pronoun gender, scope of a modifier, whether a word is a noun or a verb. A reasoning model can explicitly weigh the possible readings before committing, rather than a standard MT model’s single forward pass that picks the statistically most common mapping and moves on.

Handling low-context or terse source text. Documents, subtitles, and UI strings often arrive with little surrounding context. A reasoning approach lets the model consider “what is this sentence likely part of” before translating, which conventional NMT systems structurally can’t do since they translate segment by segment.

Style, register, and terminology consistency. Reasoning models can hold instructions like “use formal register” or “keep this term consistent with the glossary” as an explicit constraint to check against, rather than trusting it to be implicitly learned from training data.

Idioms and culturally-bound expressions. These require recognizing that a literal translation would be wrong and reasoning toward what the expression actually means before re-expressing it, which is closer to how a human translator works than to statistical pattern matching.

Self-correction on obviously wrong output. A model that can check its own draft translation against the source (e.g., “did I preserve the negation? did I drop a clause?”) catches a class of errors that standard NMT and even non-reasoning LLM translation miss.

High-quality step-by-step training data — problems paired with detailed intermediate reasoning traces, not just question-answer pairs. This is often synthetically generated or curated from expert solutions (code with debugging traces, workflow details that mimic an actual human translation-and-revision process, translator-to-reviewer conversations)

Verifiable or checkable answers — reasoning models trained via reinforcement learning benefit enormously from tasks where correctness can be automatically verified (math with a known answer, code that passes or fails unit tests, compliance with a style guide), since this gives a clean reward signal

Diverse problem difficulty — a spread from easy to very hard problems helps the model learn when extended reasoning is actually needed versus when it’s wasted effort

Clean, well-structured domain data at inference time — to perform well on a specific task, the model still needs the underlying facts, context, or documents to be accurate and complete; reasoning helps the model use information well, but it doesn’t invent missing information reliably

Sufficient context relevant to the specific query — since reasoning models are effectively doing more work with what’s given, poor or incomplete input context tends to get amplified rather than compensated for

Optimal Language AI use therefore involves moving beyond the common enterprise question “Which model translates best?” toward a more granular question:

“Which output is preferred by qualified reviewers for this content, audience, locale, channel, and use case—and can we convert that judgment into reusable training and evaluation signal?”

That has several direct implications.

• Fluency is not quality. A fluent output can still use the wrong register, address the wrong implied audience, over-explain, retain source-language discourse patterns, or violate local writing norms.
• Semantic fidelity is necessary but insufficient.The model must carry forward intent, audience relationship, genre conventions, usability, and local communicative expectations.
• Human edits are high-value data. A reviewer’s replacement of an impersonal construction with a direct-address construction contains a compact but rich expert decision. It encodes a target-language preference and the reason it applies.
• The unit of improvement is not always a full translation. Edits, ranked alternatives, annotations, rationales, error categories, and context-rich preference pairs may be more useful than undifferentiated bilingual parallel text.
• Translation should be optimized by task. A public leaflet, a regulated product instruction, a marketing page, customer-support guidance, in-product UI, and a conversational speech interaction should not share one undifferentiated definition of “good.”

Improvement will increasingly be driven by a feedback loop between models and qualified human expertise. We have seen that in some of the most successful language model implementations in the frontier science and discovery arena (AlphaFold, Molecule design, Viruses, Space travel) that there is a significant involvement of expert human steering and feedback. In frontier science uses of these models, progress is shaped by expert judgment. A specialist who determined what a correct answer looks like also needs to be captured and used to steer model preferences.

1. Models generate multiple viable alternatives

Modern multilingual LLMs and MT systems will increasingly produce translations that are broadly intelligible and often fluent. The important bottleneck is no longer always producing a grammatical candidate; it is reliably selecting or generating the most appropriate candidate for a specified target situation.

2. Experts supply preference and rationale data

The key data asset is not simply “source sentence → corrected target sentence.” It is closer to:

Data element	What it teaches the system
Source content	What must be preserved or conveyed
Candidate output	What the model or prior workflow produced
Preferred output	What a qualified reviewer judges better
Context	Audience, genre, channel, locale, reading level, brand, risk level
Rationale	Why the preferred version better serves the reader
Error or preference type	Register, terminology, omission, redundancy, tone, cultural adaptation, clarity, factuality, and so on

It places translation improvement squarely in the modern alignment-data paradigm: models improve when high-quality human judgment teaches them which of multiple plausible answers is preferable in context.

3. The data becomes training, routing, and evaluation infrastructure

The same expert signals can serve three purposes:

• Model adaptation: Fine-tuning, preference optimization, reranking, prompt design, or retrieval augmentation.
• Intelligent workflow routing: Identifying when a model can proceed automatically, when it needs terminology or reference material, and when a linguist must review.
• Evaluation: Building realistic test sets that measure the distinctions linguists actually make, rather than only generic textual similarity.

The strategic resource is not only the TMS, LLM, MT engine, or even the volume of bilingual content. It is the organization’s ability to capture contextualized human judgment systematically. This means that post-editing should be treated as model development work.

Post-editing should not be managed only as a cost to minimize or as a binary pass/fail control. It can generate reusable data—provided the operation preserves the input, the alternative, the final preferred wording, relevant context, and, where practical, the reason for the change.

For example, instead of recording only the basic source and target text in the following example:

• Source: “Applicants must submit proof of residence.”
• Final French: “Vous devez fournir une preuve de résidence.”

capture a structured signal such as:

• Content type: public-information leaflet.
• Intended reader: applicant/resident.
• Target reading level: Grade 6.
• Preference: direct second-person address.
• Reason: direct address is clearer and avoids redundant reference to “applicants.”
• Quality dimension: audience/register/conciseness.

The structured signal is substantially more useful for future AI behavior than a sentence pair alone.

The strongest implication of this trend is that language technology will progress through a shift from general capability to context-sensitive preference learning.

The practical conclusion is not “replace translators with LLMs.” It is closer to: use LLMs to generate and scale candidate language work, while turning linguists’ highest-value contextual decisions into the
assets that make the AI genuinely better.

That is especially relevant for enterprise localization. The organizations best positioned to benefit will be those that can convert linguistic expertise from an invisible service layer into a governed data, evaluation, and continuous-improvement system. The models may become increasingly capable and multimodal, but the differentiator will be who possesses the best evidence of what “right” looks like for particular audiences, languages, and high-value tasks.

Lara 3 is an example of this new type of reasoning model that learns from experience, and that is designed to gather and use an increasing array of expert human preference signals.

It is a system that is designed from the outset to learn from expert human judgments and encode these human preference signals as machine intelligence.

AMTA’s biannual conference, held in Québec City earlier this month, drew fewer attendees than in previous years — a pattern seen across many language-industry events — yet the presentations were substantive and the discussions unusually focused. Several themes that emerged across sessions directly echo the shift described above: the limits of scaling alone, the persistent difficulty of true quality evaluation, and the growing recognition that expert human signals must become first-class training and evaluation assets.

Some of the highlights that could be gathered across multiple sessions include:

• While LLMs are improving significantly in coding, agentic services, and content generation capabilities, the progress in multilingual and cross-lingual capabilities is lackluster as new models emerge.
• The transition from NMT to LLM MT (a majority of the people present were well along in their transition) with very specific proof of the improvements in quality.
• While there were several discussions on MTQE, there were honest admissions that it is often inadequate, incorrect, and unable to catch “false fluency”. Apparently, less than 3% of Fortune 500 companies use MTQE in production settings, but the growth has been 3X in the last 12 months. The TMS-based QE capabilities apparently have 0% adoption within the Fortune 500!
• There were several specific examples of how smaller LLMs were outperforming the large foundation models on translation-focused tasks.
• A fascinating though technically dense presentation on how to improve tokenization and cross-language performance by the “Best Thesis” award winner Elizabeth Salesky of Google DeepMind.
• Quality evaluation challenges included better handling of “false fluency”, LLM-as-a-Judge failures, and an urgent need for “better data and evaluation” to get a more accurate sense of model performance. Many of the people who think that “Translation is a solved problem” are using tests that are too easy
  or have been gamed. Most experts admit that Translation remains a major challenge that requires continued investment and research.

In his keynote speech, Paco Guzman of Handshake AI (formerly with major Meta MT initiatives) countered the misconception that “translation is solved,” presenting a practical perspective grounded in industry expertise.

He articulated a compelling case for driving future advancements through the collection and application of expert human judgments. According to Guzman, AI training is undergoing a paradigm shift from static datasets toward dynamic reinforcement learning environments.

Consequently, the critical constraint is no longer the size of the bilingual corpus, but the quality of the environment and its scoring mechanism. This evolution shifts translation production toward what he termed “knowledge localization”, that is, by focusing on accurately conveying concepts within context rather than merely performing word-for-word linguistic mapping.

According to his account, achieving this relies on model-agnostic harnesses paired with rubrics as the new standard of reference.

A rubric is a structured set of explicit criteria used to assess the quality of an AI system’s outputs. It breaks down otherwise subjective judgments (e.g., “Was the response good?”) into clear, reusable dimensions such as accuracy/factuality, relevance, completeness, helpfulness, safety/harmlessness, style/tone, grounding in sources, or task-specific requirements.

The rubric is what generates the reward signal during reinforcement learning training. If a task has a clean, verifiable rubric, it’s much easier to train a reasoning model to do well on it, because the training process gets an unambiguous signal of success or failure at scale.

Rubrics typically include scoring scales (e.g., 1–5) with descriptions for each level and can be applied by human reviewers or by an “LLM-as-a-judge.”

Well-designed rubrics make evaluation more consistent, reproducible, and diagnostic by highlighting why an output succeeded or failed rather than just giving a single score.

A harness is the surrounding system that runs a model through a task and manages the interaction — it’s the scaffolding, not the model itself. It typically handles things like:

• Presenting the task or prompt to the model
• Giving the model access to tools if the task requires them
• Managing multi-step interactions (letting the model take several turns, see intermediate results, and continue)
• Collecting the model’s final output in a structured, checkable format
• Enforcing constraints like time limits or step limits

In short, the harness is the automated “test suite” that turns evaluation from ad-hoc manual review into a repeatable, scalable process.

Together, rubrics define what “good” looks like, while the harness provides the machinery to measure it reliably at scale.

The harness runs the task and produces an output. The rubric judges that output. Together they’re what turns “let the model attempt lots of variations” into a usable training or evaluation signal — the harness makes the attempt possible and consistent, and the rubric decides whether the attempt succeeded.