Who Gets the Credit When a Weak Model Wins? The Truth About AI Benchmark Scores Is Even Less Reliable Than You Think
Author: Robonaissance Research Team
Compiled by: TechFlow
TechFlow Summary:In the increasingly fierce benchmark race among major mainstream large language models, the statement "Model A scores 8 points higher than Model B" is often crudely interpreted by the tech industry and capital markets as a clear difference in intelligence. However, the latest empirical research from Google Cloud AI Research and Northwestern University completely shatters this myth: without touching any underlying model weights, simply reconstructing the external workflow harness can allow previously lagging models to easily surpass top-tier models in office and professional tasks.
This discovery reveals a brutally overlooked reality: the scores and capability performance of large models in real-world scenarios have never been determined solely by the underlying models themselves, but are comprehensively driven by an environmental system co-constructed from system prompts, memory, skills, tools, and orchestration scripts. The so-called benchmark leaderboards are essentially not intellectual duels between two naked brains, but rather a performance comparison of two people working in offices with different configurations. Understanding this holds strong warning value for the current industry ecosystem that blindly chases LLM benchmarks while ignoring application engineering architectures.
An AI agent was assigned a programming task and displayed an extremely diligent demeanor.
It painstakingly explained which code file it intended to open, which terminal command it was preparing to run, and what validation checks it planned to perform.
However, the real fatal flaw lay in its mere "lip service." The agent merely listed all tool calls in plain text, resembling an apprentice standing aside with hands in pockets, verbally guiding repairs without actually writing a single line of code. Ultimately, the task scored only 60 out of a possible 100.
Subsequently, the researchers modified only one configuration file in the agent's environment. It was a brief guide for tool invocation, explicitly instructing it to stop outputting fake plain-text call commands. The underlying model was identical, its neural network weights untouched, and the test task unchanged, yet the final score skyrocketed to a perfect 100.
This vivid case comes from a cutting-edge academic paper jointly published by researchers from Google Cloud AI Research and Northwestern University on October 9, 2026. The testing data in the paper are based entirely on the authors' own experimental records and have not yet undergone fully independent external replication.
This may seem like a minor, slightly humorous interlude, but it points directly to a profound core proposition: whenever we read promotional slogans claiming "AI Model X defeated Model Y in tests," our minds typically conjure images of two pure brains competing head-to-head.
However, what is actually compared in reality are two brains each placed within specific system instructions, dedicated auxiliary tools, and work habits, with vastly different amounts of engineering optimization effort poured into each system behind the scenes. The final evaluation score always belongs to the organic combination of these three factors, and can never be claimed solely by the model itself.
Today, this peripheral harness optimization work around models can already be handed over to another AI for automation. In these practical tasks, the degree of tuning applied to this outer environment is sufficient to completely overturn the original ranking of model superiority. Therefore, when faced with any claim that "Model A crushes Model B," a truly insightful question should be: In what set of system environments does it operate? And how much auxiliary support does it actually receive?
Comeback and Overtaking: How Weaker Models Defeat Stronger Ones
The most shocking research conclusion of this paper lies in revealing a dramatic reshuffling of rankings.
In a suite of 100 office benchmark tests covering coding, web search, and document processing, each task was scored out of 100 using hidden test cases, file integrity verification, and an LLM judge. In a standard out-of-the-box default system environment, the domestic open-source model GLM-5 averaged a test score of 64.2; in the exact same standard default environment, Google's Gemini 3.5 Flash averaged a high score of 72.4. By the metrics of this evaluation benchmark itself, Gemini is undoubtedly the stronger contender, leading by roughly 8 points.
Yet, after only a single round of automated fine-tuning to its surrounding environment configuration, GLM-5's average score directly jumped to 75.7. This was not only a full 11.5 points higher than its initial score, but also 3.3 points higher than the out-of-the-box Gemini.
Under a second ready-to-use out-of-the-box test configuration, the same pattern repeated: the optimized GLM-5 scored 75.3, overtaking Gemini's 74.5. The paper's authors concluded straightforwardly: a relatively weaker model, equipped with a superior system harness, can absolutely defeat stronger models trapped in their native, rudimentary environments.
This statement is true, but its actual meaning is far more rigorous and restrained than it sounds. It absolutely does not imply that GLM-5's underlying cognitive capabilities have surpassed those of Gemini. Because if Gemini were subjected to the exact same system-level tuning, its score would also rise accordingly, climbing to 77.5.
What this experimental result truly proves is: in this category of office tasks, the generational gap in native capabilities between the two models is far smaller than the score chasm that can be opened by a single round of external systems engineering tuning. Such a degree of ranking reversal doesn't even require operating on the weights of either model itself.
Dissecting Model Harnesses: Office Components Beyond the Brain
In the AI engineering community, this layer of system environment surrounding the models is vividly called the "harness," though its essence is actually far more simple and straightforward than the technical term implies.
Imagine what an office would provide a new employee on their first day, besides bringing a smart head: a written job description, a list of rules and regulations, handover notes left by the previous employee, a thick standardized operating procedure manual, software operation guides, several ready-made scripts for automating tedious mechanical tasks, and a team of colleagues ready to collaborate and divide responsibilities at any time.
The AI harness is fundamentally the digital twin of this work package. The paper systematically dissects it into eight core components: Standing Instructions, Guardrails, Memory, Skills, Tool Guidance, Scripts, Helper Agents, and a lightweight plugin capable of dynamically intercepting and correcting agent behavior. Each of these is an independent file, and none alter the underlying model weights. They only change what the model is told, what it remembers, and what it is allowed or reminded to do.
This workplace analogy fits perfectly on another dimension as well: placing the same exceptionally talented individual in either a highly organized, efficient office or a chaotic, disorganized management quagmire will likely yield two completely different people in your evaluation. No one would consider such a test a pure measure of that employee's personal ability; yet today, the industry almost invariably falls into this unidimensional cognitive trap when interpreting "Model A scored 72 points."
Adaptive Evolution Under Non-Standard Answers: Relying on Behavioral Correction Rather Than Right/Wrong Scoring
The most intuitive way to optimize a harness is naturally to continuously attempt modifying configurations and retain changes that raise scores. But this approach presupposes an existing scoring rubric. When facing brand-new open-ended tasks in the real world, agents have no preset scoring standards because no one can provide a single definitive answer in advance. The innovative method proposed in this paper cleverly bypasses this barrier by comparing the agent's "behavioral patterns" rather than simply judging the "absolute rightness or wrongness" of the results.
The step-by-step optimization mechanism works roughly as follows: the agent first runs the task completely once under the current harness environment, and this execution log serves as the baseline reference sample (Reference Run). Next, a model responsible for editing and tuning reviews the entire execution trajectory and proposes two sets of optimized, brand-new harness plans: the first specifically targets the agent's universal behavioral weaknesses, such as deficiencies in task planning, tool invocation, and context tracking; the second focuses on supplementing domain knowledge and hard business rules for that specific vertical sector.
Then, these two modified harnesses are run again on the exact same task, and their execution processes are carefully compared against the original baseline sample: Did this modification precisely correct the anticipated defective behavior, prove completely ineffective, or accidentally disrupt other normal steps? As the final step, the system filters and retains the improvements proven effective through actual practice, merges them into a final harness, and uses it to generate the ultimate deliverable. After the task concludes, the system archives this accumulated valuable experience into a case library for future similar tasks to draw upon as needed.
The official description of this method is extremely candid regarding its role in baseline runs: it "reveals the behavioral characteristics of the harness, not whether the final deliverable is absolutely correct." This is the honest core of the method. It cannot a priori distinguish whether an answer is right or wrong; what it can observe and capture are flaws in the execution behavior itself, and what ultimately gets successfully repaired are precisely those visibly apparent behavioral shortcomings.
Let's examine some typical failure cases successfully repaired by this method: In a complex professional office task, the agent needed to synthesize a briefing memo from multiple scattered data source files including schedules, activity venues, personnel hours, and financial budgets. Its initial run scored only 43.1. Execution logs showed that during the entire process, the agent never cross-checked numbers across different files, nor did it reopen the generated document after drafting. The auto-tuned harness later forced it to list every input basis item by item, trace the original source of every number, and finally run a script to reopen the finished document for review, causing the final task score to jump directly to 77.8.
In another legal practice task, the agent needed to review and annotate an arbitration procedural order. Initially, it saved all modifications within the Word document's track-changes feature, meaning anyone reading the file in plain text would only see the final text with revisions accepted, leaving all annotation traces completely invisible. A newly injected skill pack forced it to explicitly insert addition and deletion markers directly into the main body of the text, causing the score to instantly skyrocket from 44.4 to 94.5. In the third case, an assistant agent tasked with gathering information became trapped in an endless retrieval loop, remaining empty-handed with no deliverables generated as the submission deadline approached; a guardrail plugin that issued a mandatory alert after detecting a predetermined number of pure tool invocations completely resolved this chronic procrastination problem.
It is obvious that none of these defects stem from a lack of knowledge reserves in the models. These agents possessed ample knowledge of legal statutes, arithmetic logic, and coding standards; what they truly lacked was simply good professional habits: rigorously reviewing your work product, putting findings down in black and white, ensuring readers can clearly see your modifications, and decisively hitting the execute button when necessary. Professional habits are clearly visible in execution trajectories, which differs entirely from abstract answer correctness. This is precisely why comparing execution trajectories allows us to pinpoint the root cause of problems even without standard answers. This also reasonably explains why in two out of the three benchmark tests, the relatively weaker models actually achieved larger score increases, as they inherently had more bad habits to break.
The Cost Behind the Overtaking: Stronger Editors and Additional Compute Consumption
It must be pointed out that the optimized GLM-5 did not win this contest fighting alone. Its underlying harness was personally rewritten by Claude Opus 4.8, whose capability tier significantly surpasses any competitor in this matchup; furthermore, before producing the final deliverables, each task was repeatedly run multiple times: once as a baseline, twice for two candidate solutions, and once more for the merged final version. In contrast, the Gemini in the control group ran only once under its standard default harness.
Therefore, the essence of this comeback and overtaking is GLM-5 stacking a more powerful editorial mentor and several-fold additional runtime overhead to face off against naked, unassisted Gemini. The resources and compute costs consumed by both parties are completely asymmetrical, and the experimental results themselves have never concealed this fact.
However, two pieces of powerful empirical evidence block the tendency to simply attribute this achievement to "relying entirely on a stronger model to manufacture it for it." First, researchers specifically designed a set of ablation studies where models self-edit and optimize their own harnesses, covering five models ranging in tier from GLM-5 to Claude Opus 5. Test results show that every model achieved a comprehensive uplift in its own scores across all benchmarks under this mechanism. This means the core logic of this method does not absolutely depend on an externally stronger mentor model, although the main experiment data table still employed the strongest model as the editor endpoint to pursue optimal results.
Second, according to the authors' meticulous compute accounting ledger (combining all tokens consumed during the tuning/editing process with those consumed during actual task execution), within the compute budget testing ranges they established, this method still outperforms all comparable peer schemes in terms of performance score output per unit of compute cost spent. This is a rigorous empirical finding within a specific budget range, rather than proof that it represents the absolute best paradigm regardless of infinite compute investment.
Equally important to heed is that modifying harnesses by no means naturally guarantees positive gains. Among two other agent enhancement schemes already published in the industry, scores sometimes fall significantly below those of the unmodified original harness, with the maximum drop in programming tasks reaching up to 6.6 points; another scheme saw its scores plunge below the initial baseline in no fewer than 7 out of 12 benchmark-and-environment combination tests. Even using this paper's own editor, blindly rewriting the harness without closed-loop verification via baseline runs and case libraries could cause programming scores to cliff-drop from 73.4 to 63.2. This layer of system architecture surrounding large models possesses a highly leveraged effect, sufficient to cause drastic fluctuations in final performance; the reason this experiment managed to push it steadily upward is entirely attributable to strictly adhering to one iron rule: subjecting every modification to rigorous verification within the agent's actual runtime behavior.
Boundaries and Insights of Empirical Conclusions: Benchmarks Are Never Naked Brain Duels
Based on the solid boundaries supported by existing data, we can reach this clear judgment: in office, professional consultation, and legal practice tasks, the volume of the harness surrounding large models is sufficiently massive to completely reshape the relative win-loss rankings between models; simultaneously, its adjustment cost is sufficiently low to fully enable dynamic customization for each specific task. This is precisely the entirety of what the paper robustly proves.
But it absolutely does not prove that these improvements represent an absolute intellectual leap in a deeper philosophical sense. The evaluation of these tasks relies heavily on LLM judge scoring, and the criteria in legal testing are based on quantified point deductions rather than perfect completion of the entire task. The authors themselves acknowledge that purely comparing behavioral patterns without a definite factual baseline cannot mathematically guarantee absolute correctness, nor can it ensure that a specific modification will seamlessly generalize and transfer to entirely new scenarios. Furthermore, the test set scale is relatively limited (each domain contains only 25 to 100 tasks), and the occasional fluctuation of a tiny fraction of tasks can skew the overall average; finally, the data here does not showcase miracles of continuous snowball-style self-evolution. The research team conducted no ultra-long-cycle workflow stress tests, nor have they ever claimed to have achieved boundless leaps in self-intelligence.
However, despite the aforementioned objective limitations, the technological evolution direction it indicates is incredibly clear. If a model's system environment can be automatically optimized and reconstructed by another AI within just a few rounds of task execution, then the so-called "naked base model under a default harness" ceases to constitute a natural benchmark unit for measuring AI capability. It simply becomes one ordinary option among countless possible configurations, and is very rarely the exact architecture anyone would genuinely adopt in real-world production practices.
We can absolutely not expect waiting for a more perfect foundational LLM to make this layer of engineering harness vanish into thin air. The vast majority of bugs repaired in this experiment were tightly interwoven with specific real-world work environments, such as underlying revision marks invisible to readers, or investigative tasks that drag on until the delivery deadline without a single word drafted; every brand-new workplace environment in reality brings its own unique friction. Even the strongest model evaluated in this paper still gained several points after undergoing its own system environment tuning. A more powerful model might require fewer reminders and corrections, but it ultimately still needs to clock in at a specific office, and the operational manuals there must certainly be written in advance by certain people or certain agents.
The agent that initially only verbally recited tool commands on the screen possessed an inner cognitive capacity no less than half that of the agent who later truly executed them. They are completely, utterly the same brain; the only difference is a missing key reminder in the operating manual. Next time you see a benchmark leaderboard on social media declaring that one large model has absolutely conquered another, please remain soberly aware: that leaderboard merely records the report cards submitted by two brains working in offices with vastly different configurations, after undergoing a certain number of trial-and-error attempts. Until you thoroughly figure out who decorated this office behind the scenes for them, and exactly how many retest opportunities each was granted, the genuine intelligence information you can glean from that chart is far more narrow than it appears.